JP5041041B2 - Image classification device - Google Patents

Description

  The present invention relates to an image classification device.

  2. Description of the Related Art Conventionally, there is a technical field called similar image search that searches for a similar image for one model image presented by a user. Non-Patent Document 1 uses a color histogram to roughly quantize those bins by uniformly integrating them, and use the values themselves as feature quantities to extract similar images by measuring the distance of similarity in the feature quantity space. Techniques to do this are disclosed. Non-Patent Document 2 proposes a system that searches for similar images from the respective sides of color, texture, and shape, and features similar to non-patent document 1 for color and completely different for other aspects. A feature value is defined. Non-Patent Document 3 discloses a method for retrieving similar images based on texture feature amounts. Here, the image is subjected to Gabor wavelet transform, and a set of the average value and standard deviation of the values of each generated high-frequency subband is used as a feature vector. Then, a technique for extracting an image similar to the texture shown in Brodatz texture database by distance comparison in the feature amount space is disclosed.

  On the other hand, unlike similar image search, Patent Literature 1 discloses a technique that can also be called a sensitivity search that classifies a photograph by a sensitivity adjective. Here, the photograph is approximated to three representative colors, and collated with a database that describes the relationship between a three-color color model and impression words using adjectives prepared in advance for color designers who create clothing, interior, and cityscapes. This describes the sensitivity of the photo. That is, instead of describing the method of Non-Patent Document 1 more roughly and determining a representative color, a plurality of 1 to 10 pattern models are prepared for one term.

  Further, Non-Patent Document 4 clarifies the relationship between the image and glossiness. That is, it is pointed out that there is a deep relationship between the asymmetry of the luminance histogram of an image and the mechanism by which humans perceive glossiness. Specifically, the relationship between the skewness of the luminance histogram and the glossiness is clarified. In order to create a simulation image for performing a psychological experiment for this purpose, a psychological experiment is performed by assuming a beta function that is easily correlated with the skewness as a histogram model and changing the parameters.

Japanese Patent No. 3020887

Y. Gong, C.H. Chuan, and G. Xiaoyi, "Image Indexing and Retrieval Based on Color Histograms," Multimedia Tools and Applications 2, 133-156 (1996). W. Niblack, R. Barber, W. Equitz, M. Flickner, E. Glasman, D. Petkovic, P. Ynaker, C. Faloutsos, and G. Taubin, "The QBIC Project: Querying Images By Content Using Color, Texure , and Shape, "SPIE Vol. 1908, 173-187 (1993). B. S. Manjunath and W. Y. Ma, "Texuter Features for Browsing and Retrieval of Image Data," IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol.18, No.8, August 1996. I. Motoyoshi, S. Nishida, L. Sharan, and E. H. Adelson, "Image statistics and the perception of surface qualities," Nature, 2007, May 10; Vol.447 (7141), pp.206-209.

  The methods of Non-Patent Documents 1 to 3 described above have the ability to collect similar images of scenes that match the color histograms and texture patterns of the presented images or the presented patterns very accurately, but the same sensitivity is also used for scenes of different colors and textures. There is a problem that it does not have the ability to capture the features common to images that evoke the image. On the other hand, even if the concept of the representative three-color approximation in Patent Document 1 is valid in a certain aspect, it cannot be said that the entire sensitivity of a photograph is accurately described. Non-patent document 4 is extremely remarkable, but only points out the relationship between the degree of distortion and glossiness, and the relationship between various sensibilities is completely unknown.

  An object of the present invention is to provide an image classification device that classifies images by adjectives. In addition, by introducing a hypothesis about the mechanism of sensitivity and mathematically modeling it, the relationship between the image observation amount and the psychophysical amount is investigated, and a physical quantity that characterizes the sensitivity in a sophisticated form that more closely matches the sensitivity mechanism. The purpose is to introduce a quantitative description method.

An image classification apparatus according to the present invention includes an image input unit that inputs an image, a distribution function generation unit that generates a distribution function of pixel values of the image, and the distribution function that forms a complete system and has different weights within a distribution range. Characteristic features of the shape of the distribution function by calculating a series using orthogonal basis functions and describing an index including at least the sum of the quadratic form of the expansion coefficient by describing means for describing the distribution function by an expansion coefficient And by rearranging the index in order of the magnitude of the value, the image is divided into at least two categories of images, the image having the property of the index and the image having the opposite property. And a classification means for classifying.
An image classification apparatus according to the present invention includes an image input unit that inputs an image, an edge image creation unit that creates an edge image related to an edge component of the image, and a distribution function creation unit that creates a distribution function of pixel values of the edge image The distribution function is series-expanded using basis functions orthogonal to each other with different weights in a distribution area, and the distribution function is described in terms of expansion coefficients, and a quadratic form of the expansion coefficient An evaluation unit that evaluates the feature of the shape of the distribution function by calculating an index including at least the sum; and an image having the characteristics of the index by rearranging the index in order of magnitude of the value And classifying means for classifying the images having the opposite properties into at least two categories of images.
An image classification apparatus according to the present invention includes an image input unit that inputs an image, a distribution function generation unit that generates a distribution function of pixel values of the image, and an even function group that forms a complete system and is orthogonal to each other. A series expansion using a basis function consisting of and an odd function group, the description function describing the distribution function by expansion coefficients, and the properties of the even function and the odd function belong to the same subgroup in the basis function group An even number of quantum numbers (0,2, ...) and an odd number of quantum numbers (1,3, ...) are assigned to the function groups, respectively, and the expansion coefficient and the quantum number assigned to each A physical quantity calculating means for calculating a physical quantity obtained by summing all the expansion coefficients, an evaluation means for calculating an index for evaluating the feature of the shape of the distribution function based on the physical quantity, By rearranging the images in order, And images with from that characterized in that a classifying means for classifying the images of at least two categories of images with opposite properties.

  According to the present invention, in order to extract a Kansei feature quantity from a distribution function, a solution satisfying a linear differential equation that explains many linear physical phenomena in the natural world is stretched in advance so that a more linear relationship with Kansei can be easily created. By expressing the distribution function in terms of the Hilbert space using a set of orthogonal basis functions, the expansion coefficient is linear to the human sensibility depending on the role of the weight function to form an orthogonal system according to the distribution characteristics that each distribution function can take. It is now possible to project into a space that is easy to behave on a large scale. Furthermore, the shape evaluation of the overall distortion and spread of the distribution function that is evaluated using the expansion coefficient and the information on the symmetry / asymmetric property of the basis function group and the quantum number related to the spread is also applied to human psychology. It has become possible to perform in a quantity that is easy to behave on a linear scale.

It is a graph of T ₁ (x), T ₂ (x), T ₃ (x) in the Chebyshev polynomial T _n (x). It is a graph of n = 1-5 of the basis function regarding the root of a spherical Bessel function. It is a graph of j ₀ (x), j ₁ (x), j ₂ (x) of the spherical Bessel function. It is a graph of spherical Bessel functions j ₀ (x) to j ₅ (x) defined to be expanded in the negative region. It is a block diagram which shows the image classification device which concerns on embodiment. It is a flowchart which shows the process in the image classification device which concerns on embodiment. It is a wave form diagram of synthetic waves T ₁ T ₃ and T ₂ T _{4 in} a symmetric state, which is one component to be spured with Fm = 2 (α) (α) +. It is a figure which shows the color histogram shape of the HVC surface of a "static" image and a "dynamic" image. It is a figure which shows the color histogram shape of the HVC surface of an "tight" image and an "open" image. It is a figure which shows the mode of the subband division | segmentation by 4 steps | paragraphs of wavelet transformation. When the predetermined arithmetic mean is taken for the sensitivity invariant Gz, m (α) (α) +, when the expansion order is N = 100 and the quantum number difference m is shifted, the extremes of the aligned image group distribution Is an example showing the shape of the distribution function. It is a figure which shows the color histogram shape of the HVC surface of a "striking" image and a "mysterious" image. It is a figure which shows the distribution function of a color and a texture. It is a conceptual diagram of the higher order momentum space at the time of performing a Fourier transform. It is an image figure which shows the sensitivity group in a phase space. It is a figure explaining the relationship between a position and a motion about uncertainty. It is a matrix diagram showing a mode that four expansion spurs are comprised. It is a distribution map of the energy value in the model image of an adjective. It is an imaginary figure of the hierarchical structure of a pyramid. It is an energy band figure in a color and a texture. It is an energy band figure in the conduction band of nickel.

Prior to a specific description of the embodiment of the present invention, a principle explanation will be given.
[1] Applicant's findings from previous experiments In order to address the above-mentioned problems, the applicant's previous attempts have been summarized here. Many of them are also described in Japanese Patent Application No. 2008-23469 (prior application 1) and Japanese Patent Application No. 2008-235578 (prior application 2) already filed by the present inventor.

  First, in order to prepare the experimental environment, an adjective expressing the sensitivity received from each image was assigned to about 200 images. These images were converted to Munsell HVC color space, and the correspondence between color histogram and texture PDF and additional adjectives was compared for all experimental images. Here, the texture PDF refers to a histogram of edge images. PDF is an abbreviation of Probability Density Function, and the term is used because a high frequency band is extracted from an image and its histogram is conventionally called PDF.

  Note that texture PDF is not related to one existing high-frequency sub-band, but by integrating the high-frequency sub-band extracted with multi-resolution that was originally introduced in order to mimic a cognitive mechanism that makes visual judgment in an instant. It is a histogram of the obtained integrated edge image. In the histogram, the correlation of the spatial arrangement relationship of the contrast is reflected, and various shapes different from the generalized Gaussian function taken by the normal PDF shape of one high frequency band appear.

<Importance of description of ambiguity in sensitivity classification>
The adjective describing the sensitivity of the image used for the experiment was clarified in that the average sensitivity or the typical sensitivity of the images with different hues, brightness, and saturation is evoked. Even if the hue or brightness / saturation is the same or typical, it is an ordinary phenomenon that gives a completely different sensitivity because other elements are strong. Further, from another aspect, this fact is related to the fact that adjectives have a hierarchical structure, and that only adjectives that are close to the highest concept remain as adjectives to be given to images. The concept of adjective hierarchies is consistent with facts known in fields such as neuropsychology and also shows that the description of adjective features is not as simple as when dealing with similar image retrieval. Yes.

  As a result of the applicant's investigation of the relationship between adjectives, color histograms, and texture PDFs as described above, there is a possibility that characteristics related to similar shapes that can be read vaguely from these distribution functions are directly linked to sensitivity. I came to have confidence that it was expensive. Therefore, in the sensitivity image search, it is not important to measure the similarity of the absolute shape of the histogram as in the case of the similar image search. The important point is how to understand it as an ambiguous feature to match the human sense.

  In the above-mentioned prior application 1, in particular, the color histogram points out the importance of simultaneous description of the trend of the overall shape difference pattern in the combination of the V and C planes. Points out the importance of the difference in shape. Among them, the measure tentatively taken to extract the ambiguous shape was a statistic on the kurtosis and skewness of the V-plane and C-plane histograms in the case of color histograms. One important factor is whether it is possible to judge whether one of the VC planes has a two-band structure and one of them has a one-band structure, and the shape of the axis gradually changing between them. It was. In the case of texture PDF, the asymmetry of the histogram shape was evaluated using two indices, skewness and uniquely defined eviction. In evaluating the characteristics of the same asymmetry, the evaluation of one characteristic in two dimensions by using two indices, the index sensitive to the base and the insensitive index, shows the property of the duality. It also led to the consequence that it was important to get a connection with the adjectives that it had. The duality of adjectives is that the meaning of one adjective is divided into two major classification elements that distinguish the meaning between a major major classification element and the other adjectives contained in it. It means having both at the same time.

[2] Guidelines for issues and countermeasures <Diversification and quantification of shape recognition of distribution function>
1) Problems The above-described statistical quantities such as kurtosis and skewness actually have the following problems.
a) In a single statistic, the direct link between adjectives is sparse.
b) There is no linear quantitative relationship with the psychological quantity related to sensitivity.
c) There is a limit in dealing with diversification of histogram shape recognition.

  More specifically, statistics such as kurtosis and skewness can be used to discuss the shape of asymmetry and kurtosis as parameters that represent deviations from the Gaussian distribution. Not prepared for. Therefore, unless the kurtosis, skewness, standard deviation, average value, etc. are used in combination, it is difficult to grasp the characteristics of the shape of the histogram and deal with the adjectives. Even if they are combined, for example, the kurtosis value does not have the ability to distinguish between two band structures and a uniform distribution structure. In addition, only a few adjectives can be explained.

  Even if these statistics are standardized by standard deviation from the definition, not only values close to ± 1 near zero, but also extreme values such as +20 for kurtosis, depending on the image. derive. As a result of experimentally verifying the color histogram as a trial, one index alone was not only directly linked to higher adjectives, but also far from the psychological scale.

2) Policy of countermeasures In order to solve the above-mentioned problems, a completely new concept is introduced. In other words, it tries to describe sensibilities mathematically and concisely by capturing sensibilities from a physical point of view and developing dynamics. In the following, the name of histogram is generalized and referred to as a distribution function which is a statistical physics term because it is a physical quantity that can be observed from an image. As specified in the next section, the meaning of the distribution function is not limited to the range of the histogram.

In order to diversify and quantify the shape recognition of distribution functions, we introduce quantum mechanical techniques. Here, the inferences that serve as the basis are listed.
• Photons and electrons in photographic image formation and exposure follow quantum theory.
・ The human visual system is the same, and the neural circuit in the brain is also a quantum phenomenon.

  There is a fact that sensibility fluctuates as evidence of the second content. Even if you look at an image of an external object or photo with the same field of view, the impression often varies from day to day. Considering these facts, it is considered that a linear correspondence between sensibility and image characteristics will be created if the quantum description of sensibility can be described.

[3] Descriptive form of target sensitivity <hierarchy of adjective structure>
In general, it is known that the recognition mechanism of adjectives in the brain has a pyramid-type hierarchy. In response to this fact, when performing a content-based adjective search from an image signal, what are the features related to the sensibility of the image and what structure should they have? It is natural to guess that it has a pyramid hierarchy.

  Here, the sensitivity feature quantity assumed by the present applicant and its hierarchical structure will be described. First, as a sensitivity feature amount in the lowest layer in the lowest dimension, there can be considered a scalar quantity-related element such as “representative hue, representative luminance, representative saturation” of a photographic image. In other words, if a photograph is represented by the hue, brightness, and saturation with the largest area ratio, it is an overall “red” or “green” image, or an overall “under-exposed” “dark” image. There are very low-order features that evoke the impression.

  Next, it is considered that a vector quantity element related to color described by “HVC color distribution structure” is located. That is, when the HVC histogram is in a specific characteristic state, it is estimated that a slightly higher order sensitivity is evoked. Adjectives that fall under this hierarchy are considered to be “fresh” or “slow”, for example. An image having these impressions behaves strongly as a conspicuous sensibility prior to the previous low-order impression. However, the low-order impressions do not disappear and their properties remain. It is considered that the next feature is related to “edge, texture, contrast”. At the present time, it is estimated that the vector elements of the structure factor reduced one-dimensionally with respect to the texture correspond. Depending on the number of edges and textures and the strength of the contrast, it is thought that it gives an impression that the image is a “brilliant” image or a “rough” image.

  Then, it is estimated that “composition”, that is, “spatial distribution” is located above the above. This is because “color” and “texture” described above can be discussed by extracting one-dimensional information from each color plane of HVC, but the spatial distribution must be discussed by extracting two-dimensional information from each plane. Then, it is considered that adjectives such as “noisy” and “dense” can be discriminated. It is thought that there is a structure like a universal aesthetic sense that is located above. However, it is presumed that it greatly affects the individual's sensitivity. It is estimated that an adjective such as “beautiful” can be discriminated from among them.

<Descriptive form of features unique to sensitivity: Additivity>
As a characteristic of the above-mentioned hierarchical structure, it can be said that the sensitivity at the lower level continues without disappearing, and if there is a higher sensitivity, it preferentially remains as an impression concept for the image. In addition, it can be said that it is actually possible to intentionally eliminate the higher-order concept and to give the superiority or inferiority of the features at the lower-order stage. Consideration shows that a feature amount having such a property can be described wonderfully if it has an additive property among the feature amounts. This is because the additive nature always has a mechanism that allows the addition of higher-order factors and allows replacement that surpasses the previous lower-order factors.

  Moreover, the superiority of an apparent additive property is proved by a thought experiment in the case of a feature amount having an additive property that will be described later. In other words, it is necessary to treat Euclidean geometric features independently between features with different properties, and what kinds of situations can occur with feature values with synergistic properties, unlike additive properties. By considering. Therefore, it is concluded that the affective feature quantity must have an additive property as a description form that can reflect the hierarchy of adjectives in principal axis synthesis.

  When performing linear combination between additive sensibility features, adding features with emphasis on the highest-order feature factor will also reflect the balance of the lower-order factors to some extent, while the adjective is the best. It reproduces the mechanism of human emotions, where the words of the superordinate concept are most dominantly guided. To that end, it is necessary to create Kansei feature quantities with the same quality of additivity even if the Kansei main axes are different. In other words, the affective feature quantity must be described by a physical quantity of the same dimension with additiveness. This is a condition that must be satisfied equally both within the sensitivity main axis and between the main axes.

<Sensitive linear model>
A high-order emotional feature amount that can be read from the distribution function f is defined as follows. That is, it is assumed that there are a plurality of distribution function shape features that appear in common among image groups having the same sensitivity even if the distribution function changes, for each distribution function of different types and levels. It is assumed that the number i takes values of i = 1, 2,.

Color histogram distribution F _i
Texture PDF distribution ・ ・ ・ G _i
Spatial distribution of pixel values ・ ・ ・ H _i
Assume that a linear model of sensibility is established as follows.

Adjective = α ₁ F ₁ + α ₂ F ₂ + ...
+ β ₁ G ₁ + β ₂ G ₂ + ...
+ γ ₁ H ₁ + γ ₂ H ₂ + ...
A method for constructing a feature quantity having such an additive property will be specifically described below with respect to F _i and G _{i in} the following embodiments. In the sensitivity invariants defined in the embodiments described later, the fact that the additive properties work well generates the same kind of invariants in each of hue, luminance, and saturation, and takes an arithmetic average among them. This also proves from the fact that more stable experimental results are born.

<Comparison with other non-additive methods>
1) Representation of feature vector by Euclidean distance The conventional similar image search technique handles colors, textures, and shapes individually, and compares distances of similarities in each feature space. For example, the method of Document 2 is taken as an example. The feature quantity of each axis defines a completely different type of feature quantity between the axes. Although there is no clear description of the method of synthesizing the main axes, when similar image retrieval is performed using color, texture, and shape feature quantities together, all the feature quantities of the three principal axes are made into one vector, and in the feature quantity space It is common to measure the Euclidean distance and search for images that are as close as possible in color, texture, and shape.

  If such a Euclidean distance is used, the classification index has the following properties. For example, it is assumed that a certain feature amount has no similarity in color. At that time, even if the texture and the shape are similar, if the feature amount having no similarity acts and is placed at a distance away from the Euclidean distance, it can no longer be approached. That is, there is no superiority or inferiority relationship between feature amounts, and all are treated in the same row. Therefore, the higher-order feature quantity cannot overturn the result of the lower-order feature quantity. This property is therefore incompatible with the hierarchy of adjectives.

2) Representation by synergistic feature amount Assume that each feature amount is expressed with a synergistic property. In such a case, the classification index has the following properties. It is assumed that a certain feature amount is very similar, and the feature amount model matches the value of the search target image. The difference in similarity between the feature quantities is zero, and even if there is no similarity between other feature quantities, the result of the geometric mean is zero, and finally, the similarity is extremely similar. In other words, if even one feature quantity having a synergistic property matches, all other judgments will not work. In the sense that one feature quantity is prominent, it has the property of replacement by high-order factors, but the place where other low-order features stop working does not match the property of adjectives. In other words, it is not a kind of overrule of the determination of, for example, “green image”, which is determined only by the low-order average color, and the property continues to remain no matter how low-order.

[4] Quantum mechanical description of distribution function <Hilbert space representation>
The distribution function is expanded in series by a basis function that satisfies the linear differential equation. However, these basis function systems are orthogonal to each other and have completeness in the sense that the original distribution function can be completely reproduced. It can be expressed as the following formula.

Series expansion of distribution function

Orthogonality of basis functions

  Here are the ideas behind these. In order for the sensibility feature value read from the distribution function to satisfy an additive property, first, the components of the distribution function must satisfy the linear differential equation. This linear differential equation is positioned as the equation of motion that the constituent elements satisfy. This equation of motion is based on the hypothesis that the closer to the differential equation that many physical phenomena satisfy in the world of mechanics, electromagnetism, and quantum mechanics, the closer to the physical phenomenon occurring in the brain.

  The weight function that defines the orthogonality must be determined by selecting an optimal basis function according to the characteristics of the distribution function so as to be as close as possible to the human recognition process. Functions that are orthogonal by integration are generally called special functions, and many of them are defined by hypergeometric differential equations that are linear differential equations or confluent hypergeometric differential equations. Also, in order to be able to expand series, it must have completeness in the sense that the original function can be represented equivalently. Since the special function that forms such an orthogonal system does not necessarily have completeness, the special function that has both the selection function of the weight function and the completeness is limited. The first guideline for selection is the determination of whether the basis function group is similar to the shape of the distribution function currently targeted and the consistency of the interval region.

Second order homogeneous differential equations
P (y) = y "+ p (x) y '+ q (x) y = 0
The solution y is linear for any constant C
P (Cy) = CP (y), P (C ₁ y ₁ + C ₂ y ₂ ) = C ₁ P (y ₁ ) + C ₂ P (y ₂ )
It has. Therefore, since the principle of superposition holds in the equation of motion represented by this type of linear differential equation, the general solution is represented by series expansion.

  Both hypergeometric differential equations and confluent hypergeometric differential equations are represented by the above-mentioned second-order linear differential equations, hypergeometric differential equations have definite singularities at x = 0,1, ∞, and confluent hypergeometric differential equations Has a definite singularity at x = 0 and an uncertain singularity at x = ∞. Each equation is described (see Document B2).

Hypergeometric differential equation
x (1-x) y "(x) + [c- (a + b + 1) x] y '(x) -aby (x) = 0
Confluent hypergeometric differential equation
xy "(x) + [cx] y '(x) -ay (x) = 0
The description form of series expansion of the solution of the equation of motion satisfied by the state function by the function system having orthogonality and completeness is regarded as being the same as the expression in the Hilbert space. In quantum mechanics, the description of the equation of motion shifts to a matrix form by Hilbert space representation. Therefore, the construction of a Hilbert space requires expansion by a special function with complete orthogonality (see Document B1).

  In describing the sensibility dynamically, it is considered that the basis function constituting the distribution function formulates one aspect of the sensibility that at least the differential equation selected below is satisfied as the equation of motion. That is, it is considered that the hypergeometric differential equation is satisfied in the projection surface of the color distribution function and the confluent hypergeometric differential equation is satisfied in the projection surface of the texture distribution function. These linear differential equations are generic names of differential equation types and are located in different types. On the other hand, many inclusive types of differential equations can be derived by introducing parameter placement and variable transformation. For example, a hypergeometric function that is a solution of a hypergeometric differential equation can handle a Chebyshev function, a Legendre function, or the like as a special case of parameters. Further, the confluent hypergeometric function that is the solution of the confluent hypergeometric differential equation leads to a Bessel function, a modified Bessel function, a Hermitian function, a Laguerre function, and the like as special cases. Furthermore, a spherical Bessel function and a spherical deformation Bessel function are derived by variable transformation of the Bessel differential equation (see Document B2).

  In practice, it is appropriate to use the Chebyshev function for the one-dimensional distribution function of the color and the spherical Bessel function for the one-dimensional distribution function of the texture. Therefore, if a sensuous differential equation exists in the brain, the linear differential equation is satisfied, and one of the projected surfaces of the color is a hypergeometric function describing the Chebyshev function, and another edge From the side of the confluent hypergeometric function that describes the spherical Bessel function, it can be said that the texture projection surface corresponds to the description on two different projection planes. The basis functions of the differential equations to be satisfied by the respective distribution functions play a role of simulating the coordinates of the wave type signal processing system called the brain by the Hilbert space.

[Document B1] Schiff "Quantum Mechanics" (3rd edition, 1970), Chapter 6 "Matrix Form of Quantum Mechanics"
[Reference B2] George Alfken, Basic Physics Mathematics, Volume 3 “Special Functions and Integral Equations” (2nd edition, 1970; Japanese translation 1978), Chapter 1 “Bessel functions” and Chapter 3 “Special functions”
<Hilbert space representation of each distribution function>
1) Color histogram Basis function

Orthogonal weight function

Here Chebyshev polynomials can be described analytically. Takes the values n = 0,1,2, ...

Here, FIG. 1 shows a graph of T ₁ (x), T ₂ (x), and T ₃ (x) in the Chebyshev polynomial T _n (x).
The exact orthogonal relationship including normalization conditions is given below.

  The orthogonality of the Chebyshev function places a very high density emphasis on the rising and falling parts of the distribution function. That is, when dealing with a function system that has a characteristic that suddenly occurs in a finite interval and ends suddenly like a histogram, the rising and falling parts are very important as the shape of the function. It shows that it cannot be approximated in a true sense unless it is accurately described.

2) In the case of texture PDF When the spherical Bessel function is extracted for one order, the function group scaled from the low frequency to the high frequency by the number of roots existing in the interval [0, a] is orthogonal and complete with respect to the roots. Make a system. Since a zero-order function has a peak at the origin, a single series expansion is performed using the root of the zero-order function. At that time, the distribution function of the texture PDF must be developed separately for the right section and the left section of the peak.

2-1) Single series expansion Basis function

  Here, FIG. 2 shows a graph of n = 1-5.

Orthogonal weight function

Here, the spherical Bessel function can be described analytically.

Here, FIG. 3 shows a graph of j ₀ (x), j ₁ (x), j ₂ (x) of the spherical Bessel function.
The orthogonality relation for the exact root including the normalization conditions is given below.

  Since the spherical Bessel function is a function having a maximum intensity at the origin in the case of a zero-order function, when the distribution function having the maximum frequency at the origin is approximated, the weight of the origin is always removed and the shape of the base is emphasized.

  By the way, various Bessel functions are functions suitable for description when light or waves diffuse from the origin toward the periphery, or when there is a singular point such as a light source at the origin. The spherical Bessel function is suitable for describing the wave of the radial component in the spherical coordinate system, and its strength decreases more rapidly as the distance from the origin is larger than that of the cylindrical coordinate system. The one-dimensional distribution function of the edge image to be handled now is close to the property of a spherical Bessel function because of its rapid intensity decrease rate. This is because even if we consider the condensing process in which a photograph is taken through a lens, the light falling from the hemisphere is a spherical wave, and its radial component is described by a spherical Bessel function, especially non-cylindrical coordinates. This is because the reason for isotropicity cannot be found.

  Since the spherical Bessel function always has a peak at the origin when a zero-order function is used, the weight of the origin is removed in order to evaluate the shape of the distribution function with the property that the peak always appears at the origin as in the histogram of the edge image. Therefore, focusing on the nature of the base is a very compatible method. As a result, it plays the role of creating a foundation that creates a linear relationship with the essence of the event.

2-2) In case of double series expansion Since the spherical Bessel function is a function that usually describes the motion in the radial direction, it is usually defined in the positive region, but by expanding it to the negative region, Orthogonality also appears between functions of different orders. This is because the spherical Bessel function has the property that the even order is an even function and the odd order is an odd function.

Here, FIG. 4 shows a graph of spherical Bessel functions j ₀ (x) to j ₅ (x) defined to be enlarged in the negative region.
Orthogonality of the order of spherical Bessel functions

If the orthogonality regarding the order and the orthogonality regarding the root are considered together, the positive region and the negative region of the distribution function can be developed simultaneously. Since root expansion always has completeness, it expands into a double series of orders and roots. For the time being, there is only one odd function for extracting the asymmetric component, so root expansion using two of the zeroth-order even function and the first-order odd function is sufficient.

[5] Correspondence between expansion coefficient and dynamics <Problem raising>
The expansion coefficient means that there are many components having a waveform and a frequency corresponding to the basis function in the distribution function. Here, there is a problem that “the expansion coefficient c _i itself is suitable for the sensitivity feature quantity”.

<Trend of expansion coefficient in actual data>
The value of the coefficient c _i that actually expands the distribution function is likely to be directly influenced by the absolute shape of each distribution function. Therefore, the value varies greatly depending on the image. When the expansion coefficients of the distribution functions of images assigned with the same adjectives are compared directly, a slight correlation tendency is recognized, but a tendency to vary more than that is strong. Even if you try to take a statistical average of expansion coefficients between images with the same adjective and try to make the expansion coefficient a model corresponding to that adjective, most expansion coefficients disappear to zero, Even adjectives tend to converge to a constant that has no common meaning.

<Building correspondence with mechanics>
Here, we make a hypothesis that associates the following dynamics with the image system.

Dynamical system:
“Even if the motion state (momentum _pi ) of each particle changes, there is a conserved amount (energy E) that characterizes the entire motion system.”
Image system:
“Even if the pixel value distribution (state component c _i ) of each image changes, there should be an invariant (I) that characterizes the sensitivity of the entire image.”

  When a person perceives the signal distribution of an image or an image reflected in the visual field of the outside world, it feels something like a “field” energy of a certain sensibility, and instantly converts it to that kind of field energy. This hypothesis includes as a model the mechanism by which the corresponding neural state in the brain is excited and an adjective is recalled.

  The sum of the expansion coefficients in the quadratic form means that even if the distribution function changes variously, it is possible to extract the common features among the distribution functions of the image groups that recall the same sensitivity. This summing operation alleviates the requirement for strict similarity between individual elements and aggregates a group of generally similar features, thereby deriving a function that collates features that are vaguely similar in shape.

<Secondary Form and Additivity>
The reason for the conclusion that Kansei invariants must take the quadratic form is that the theoretical background in the process of deriving the equation of motion from the action function based on the principle of minimum action in theoretical physics and the process of building the Hamiltonian Is deeply involved. That is, assuming that one aspect of the kinematic field equation of motion satisfies the linear differential equation, the Lagrangian, which is the integrand of the action function, has a quadratic expression for the sensitivity field. Because it is required. This is because the equation of motion is derived based on the variational principle by putting the first derivative of the action integral equal to zero, and the superposition principle is maintained by dropping one order at that time. Yes (see Document A1). The mechanical invariant that is the first integral of the equation of motion is called the integral of the motion, and keeps a constant value during the motion. If this is sensitivity, it is considered that it corresponds to continuing to give the same sensitivity even if the signal distribution of the image changes.

  There are energy, momentum, and angular momentum in the integral of motion, and it is clarified by mechanics that these all have an important property of additivity (see Document A2). In addition, the signal distribution of an image whose sensitivity is a problem is an image having a number of pixels on the order of 10 6 to 8 and a statistical group consisting of hundreds and thousands of images. It is necessary to use statistical physics.

According to statistical physics, in the integration of additive motion, the conclusion is that after statistical averaging, only the energy determines the statistical properties of the entire system, that is, the statistical distribution of the system. The momentum and angular momentum are simply attributed to the uniform translational motion and uniform rotational motion of the entire system, and are not useful for describing the system (see Document A3). This is because the expansion coefficient c _i corresponding to the momentum derived from the distribution function to describe the sensibility tends to converge to a meaningless constant when statistical averaging is performed between image groups of the same sensibility. Matches.

When the correspondence between the distribution function f and the mechanical invariant, that is, the physical quantity of the quadratic form having the energy dimension is summarized as above,
Dynamical system Image system
f = f (E) ⇔ I = I (f)
Here, the sensitivity invariant is derived from the distribution function that is the observed quantity, so that the expression is opposite to that of the dynamical system. In addition, the Kansei invariant of the secondary form can handle various sensibilities by extending the definition as explained in the next section.

  In this way, all sensibility invariants are additive, satisfy the additivity requirement of the sensibility linear model, and the color, texture, and composition can be handled in a common manner. That is, all feature quantities having different sensitivity principal axes can be handled additively. The sensitivity invariants derived from the composition are also constructed in the same manner based on the same guidelines.

  This is because it has been determined that the above-described theoretical background is inherent in the process where the sensitivity invariant needs to be in the quadratic form as described above after experimental trial and error. In other words, even if experimentally creating many possible indicators such as the absolute value and ratio of the expansion coefficient and trying them, the image alignment that matches the sensibility with the sensitivity is not obtained many times. It is based on the fact that alignment with high sensitivity and alignment is obtained for the first time by the expression of the sum of the quadratic form repeatedly failing.

[Literature A1] Landau-Lifschitz Theory of Physics, Volume 2, “Theory of Field Classical” (6th edition, 1973), Chapter 4, “Equation for Fields”, Section 27 “Functions of Electromagnetic Fields”
[Literature A2] Landau-Lifschitz School of Theoretical Physics Vol. 1 “Mechanics” (Amendment 3rd Edition, 1973), Chapter 1 “Equation of Motion” and Chapter 2 “Conservation Law”
[Reference A3] Landau-Lifschitz Theoretical Physics, Vol. 5, "Statistical Physics Part 1" (3rd edition, 1976), Chapter 1, "Basic Principles of Statistics", Section 4, "Role of Energy"
[6] Quadratic expression of Kansei invariant <Composition of two ground states>
In order to create a quadratic form sensitivity invariant, returning to the viewpoint of recognizing the shape of the distribution function, the shape of the distribution function f is extracted by a synthesis system of the two ground states ψ _i and ψ _k . In general, it is assumed that an arbitrary function f ^(α) is expressed by expansion by n basis functions.

f ^(α) = c ₁ ^(α) ψ ₁ ^(α) + c ₂ ^(α) ψ ₂ ^(α) + ... + c _n ^(α) ψ _n ^(α)
According to group theory, n * n ground states ψ _i ψ _k represented by the direct product of a composite system are reducible expressions, and n (n + 1) / 2 symmetric products and n (n-1) It can be divided into two ground states represented by / 2 antisymmetric products (see Document A4).

  When the distribution functions of the hue plane (H), luminance plane (V), and saturation plane (C) are expanded in series using the same basis function, the characteristics of the shape of the distribution function in the same color plane are The shape of the distribution function between the color planes can be measured by the waveform expansion component created by the basis function, and the waveform expansion generated by the basis function of the antisymmetric product is combined with the waveform expansion component generated by the basis function of the symmetric product. It can also be measured by ingredients. Since the expansion components due to the waveforms of these synthesis systems are guaranteed to be orthogonal to each other in the ground state before synthesis, the expansion coefficients due to the ground state of the synthesis system are matrices in the form of symmetric products and antisymmetric products. It can be expressed as an element.

Kansei invariants are defined by taking a matrix spur. However, spur usually refers only to the sum of diagonal elements, but here it is a new extended type that takes the diagonal sum of non-diagonal components with only a constant quantum number difference between the row and column positions. Define a spur. As a result, the ratio of the components of the composite waveform of the two ground states that maintain a constant quantum number difference in the distribution function can be summed over all the composite waveforms that can be created. It can be detected in the taken state. Between two distribution functions with different color planes, the proportion of combinations of different ground states with a certain quantum number is present in two proportions: symmetric state and antisymmetric state of the combination. It is evaluated with the composite waveform.

Here, when c _i is taken as the Chebyshev expansion coefficient of the color histogram, I _i = F _i , and when c _i is taken as the spherical Bessel expansion coefficient of the texture PDF, I _i = G _i . However, the number i is assigned in order to the invariant elements obtained in plural, without distinguishing between symmetric and antisymmetric.

Here, the sensitivity invariant defined above will be compared with the field invariant in the electromagnetic field. According to the document A5, there are two types of invariants for the Lorentz transformation of the electric field E and the magnetic field H: a true scalar and a pseudoscalar.

Both are determined by a quadratic spur of the four-dimensional electromagnetic field tensor F ^ik , the latter accompanied by a completely antisymmetric unit tensor e ^iklm , and the former not accompanied by it. However, the sum symbol for the tensor is omitted by the Einstein convention.

  Therefore, a sensitivity invariant that gives the same sensitivity even when moving from one image to another has a configuration very similar to the field invariant when converted from one reference system to another reference system in an electromagnetic field. . In the previous section, we discussed sensitivity invariants corresponding to kinetic energy, but there are more than two types of “sensitivity fields”, and it is more straightforward to think that field energy is transmitted. is there. In addition, the components of the electromagnetic field are four-dimensional in time and space, but it is presumed that there are at least elements of the order necessary for describing the distribution function in the sensitivity component. However, as the neural circuit of the human brain learns about the growth process, an electric signal circuit induced in the field is constructed against these sensitive fields, or the neural circuit that is excited instantaneously in energy. It is a natural idea to acquire energy levels.

[Document A4] Landau and Lifshitz, Course of Theoretical Physics, Volume 3 "Quantum Mechanics (Non-Relativistic Theory)," (Third revised edition, 1977), Chapter 12 "The Theory of Symmetry," Section 94 "Representations of groups. "
[Literature A5] Landau-Lifschitz Theory of Physics Theory Volume 2, “Classical Theory of Fields” (Original 6th edition, 1973), Chapter 3, “Charges in Fields”, Section 23 “Electromagnetic Field Tensors”, Section Section 24 “Lorentz transformation” and Section 25 “Field invariants”
[First embodiment]
(Hilbert space representation of color histogram and linear sum of sensitivity invariants)
The image classification apparatus according to the first embodiment will be described below with reference to the drawings. FIG. 5 is a block diagram illustrating an image classification apparatus according to the embodiment. Here, the image classification device is realized by the personal computer 10. The personal computer 10 is connected to a digital camera or another computer, receives image data from the digital camera or another computer, or receives image data from a memory card installed in a memory card slot, and receives a hard disk device ( (Not shown). The personal computer 10 performs image classification processing described below on the stored image data.

  The loading of the image classification program into the personal computer 10 may be performed from a storage medium such as a CD-ROM storing the program, or may be performed via the network 12 or the like. When the network 12 is used, the program read from the hard disk device 16 connected to the server 14 is loaded. The personal computer 10 includes a CPU and peripheral circuits controlled by the CPU, and performs image classification processing shown in the flowchart of FIG. 6 based on a program in which the CPU is installed.

<Processing for search target images>
1. Conversion to Munsell HVC color space (step S1 in FIG. 6)
The input image is converted into a Munsell color space with high human perceptual color uniformity. Munsell color space is a color space that is divided so that the hue H is divided by 100 degrees around the circle, the luminance V is distributed to the level of 0 to 10, and the saturation C is distributed to the level of about 0 to 25. This is a color space designed so as to satisfy an equal rate of perceiving C color difference 2 as 1 as an equivalent color difference. Among them, the region where the C value is 1 or less, the V value is 0.5 or less, and the region where 9.5 or more is defined as N (neutral hue). From the color space expressed in the RGB space to the HVC color space, mathematical conversion can be performed approximately through conversion to the XYZ space. This is achieved by introducing an expression that corrects the lack of uniform color by using the definition of L * a * b * or L * C * H *, which is one of the uniform color spaces. It has been realized.

When the input image is, for example, an image expressed in an sRGB color space with an output gamma characteristic, it is first converted to a linear gradation and then converted into an XYZ space according to the sRGB standard.

  In the first embodiment, it is assumed that a hue surface is prepared as an N (neutral) separation surface.

2. Creation of color one-dimensional distribution function (step S2 in FIG. 6)
Create a histogram for each HVC plane. The number of bins in the histogram should be set to about 200 for all of H, V, and C. At this time, the hue surface is a surface on which N separation is performed. Therefore, the area ratio classified into N is not included in the histogram of the H plane even if it is integrated with the hue circle. Normally, N is distributed randomly in the hue circle, so a random and uniform offset distribution is excluded from the histogram of the hue plane, and the original chromatic color histogram shape remains. For convenience, if the histogram value is normalized by the number of pixels, a one-dimensional distribution function representing the probability density of the pixel value is obtained. The distribution function schematically created is expressed as follows.
f (H), f (V), f (C)
3. Hilbert space representation of distribution function (step S3 in FIG. 6)
3-1. Variable transformation If the horizontal axis of the histogram is [a, b] and the vertical axis is [fa, fb], the horizontal axis is [-1,1] and the vertical axis is [-1,1]. Variable conversion is performed in the interval that fits in. For convenience only in this section, if the variable on the horizontal axis is converted from x to y and the variable on the vertical axis is converted from fx to fy, the conversion formula is as follows.

Variable transformation on the horizontal axis: y = {x- (b + a) / 2} / {(ba) / 2}
Variable conversion on the vertical axis: fy = {fx- (fb + fa) / 2} / {(fb-fa) / 2}
Since the distribution area of the histogram of the hue plane is a hue circle, a starting breakpoint is provided, and a distribution area that goes around from there and returns to the same point is set. The starting point a is not a fixed point, but a point at which the density of the distribution function is minimum is searched for each image, and a breakpoint is inserted therein.

  Depending on how to define the variable transformation on the vertical axis, the following constant factors are added to the series expansion coefficient. Assuming that the expansion coefficient corresponds to the momentum as explained in the introduction, the constant factor remaining by the statistical average of the image group depends on how the Hilbert space coordinate system of the entire image group is selected. It is equivalent to remaining as a translational movement.

3-2. Series expansion by Chebyshev polynomials The distribution range of the horizontal axis subjected to the above variable transformation is represented by x regardless of H, V, and C. Each distribution function of HVC is expanded with N-order Chebyshev functions.

The expansion coefficient c _n is obtained by the following equation using the orthogonality of the basis functions.

However, when n = 0, specially to c ₀ = c _0/2.

Here, variable conversion is introduced.

In this way, the expansion coefficient can be obtained simply as follows.

  When the number of bins in the color histogram is about 200, the expansion order may be set to about N = 50.

4). Generation of Kansei Invariant (Step S4 in FIG. 6)
The shape of the distribution function of the H, V, C histogram is extracted by a synthesis system of two Chebyshev basis functions. That is, the sensibilities that appear due to the structure taken by the HVC distribution function are extracted as feature quantities.

  The sensitivity invariant is defined by the spur of matrix elements having a certain quantum number difference for matrix elements created by the quadratic form of the expansion coefficient. Quantum difference m can be defined up to m = 0,1,2, ..., N / 2 in the symmetric state, and m = 1,2, ..., N / 2-1 in the antisymmetric state Can be defined. An example of creating quantum numbers up to 0, 1, 2 is shown. Here, with one exception, it is standardized and defined so that all invariant values fall within [-1,1]. Further, when the sum range exceeds the range of k = 0, 1,..., N−1, it is treated that k = 0 is connected in a ring shape after k = N−1.

Evaluation of basis functions in the same color plane by a synthesis system

Evaluation of basis functions between different color planes by synthesis system

  An invariant with a plus sign indicates the proportion of the waveform that is signed due to the symmetry of the composite system in the distribution function of the image, and an invariant with a minus sign is in the distribution function of the image. It shows the existence ratio of the waveform with the sign due to the antisymmetric state of the synthetic system.

If the value of the invariant is close to zero, it means that there is no component of the composite waveform, and if it is close to +1, it means that there are many components in the form of the composite waveform. If it is close to 1, it means that there are many components of the waveform obtained by inverting the sign of the composite waveform. As an example, a waveform diagram (see FIG. 7) of the symmetric composite waves T ₁ T ₃ and T ₂ T ₄ , which is one component to be spurted with Fm = 2 (α) (α) +, is shown.

  When a sensitivity invariant with a quantum number difference of zero between different color planes shows a significant value, the distribution function of one color plane and the distribution function of another color plane always have some specific shape difference. I can say that. On the other hand, when the sensitivity invariant with a quantum number difference of zero between different color planes shows a significant value, it can be said that the shapes of the distribution functions of the two color planes are very similar.

  Only Fm = 0 (α) (α) + cannot be normalized. Since the interval of the value of the distribution function is variable-converted to [-1,1], the actual value takes a value close to zero or a value from about 0.4 to about 1.5. The meaning of this value is that the higher the value, the higher the proportion expressed in the ground state in the distribution function, and the smaller the value, the more the proportion expressed in various ground states. Indicates that it is expensive. Similarly, for Fm ≠ 0 (α) (α) +, if the absolute value of the value is large, the degree of unipolar concentration on one composite waveform in the ground state of the composite system is high, and the absolute value of the value is small. For example, it is expressed in a distributed manner in many of the ground states of the synthetic system, or it means that there are not many waveform components represented by the ground state of the synthetic system.

The symmetric state and the antisymmetric state are conjugated with each other. Regarding the sensitivity invariants of the symmetric product and antisymmetric product of the same quantum number difference, when examining the degree of alignment of the image group related to the invariant, it is unclear what sensitivity is gathered near zero in the symmetric product. A group of images appears at both ends of the index distribution in the antisymmetric product. The opposite is also true, and an image group that is unknown in the anti-symmetric product index appears at both ends of the image distribution in the symmetric product index. This fact is a natural consequence of the following mathematical formula. That is, when the value of the symmetric product is zero, c _i ^(α) c _k ^(β) = −c _k ^(α) c _i ^(β) , so that the value of the anti-symmetric product c _i ^(α) c _k ^{(β )} -c _k ^(α) c _i ^(β) = 2c _i ^(α) c _k ^(β) = -2c _k ^(α) c _i ^(β), and it is easy to take the largest value of either ±.

5. Creating an adjective discrimination index (step S5 in FIG. 6)
5-1. Linear combination of Kansei invariants Creates a new index Q _i that linearly combines Kansei invariants using the additive property of Kansei invariants as an index for searching for Kansei adjectives (i) . The adjectives that can be represented by the index Q _i are not only one adjective but also a pair of adjectives that have adjectives of the opposite nature.
Q _i = α ₁ F ₁ + α ₂ F ₂ + ...
Here, the value of the linear coupling parameter α _i is normalized so that Q _i becomes an index in the range of [−1, 1] again.
5-2. Parameter setting for search target adjectives Linear combination parameters corresponding to predetermined adjectives are learned in advance, and the model parameters are set.

6). Image classification process (step S6 in FIG. 6)
The images are classified based on the adjective discrimination index. For each image in the image database group to be input, an adjective discrimination index Q _i is calculated and rearranged in order of the value of Q _i . The frequency distribution of the image group with respect to the discriminant index Q _i is shaped like a Gaussian distribution or Poisson distribution, so it is unique to the adjective discrimination index at a statistically significant level at both ends than the other image groups. An image showing sex is presented.

  A specific example of how an image that gives an actual feeling is extracted will be shown. The simplest linear combination is when the coefficient parameter is finite only for one sensitivity invariant and all others are zero. Only their respective properties are shown. A method of determining when a plurality of coefficient parameters remain will be described in the model learning section described later.

  In fact, it is projected on the Hilbert space where physical phenomena are easy to describe in this way, and continues to act invariantly on changes in the signal distribution of the image. From the above, it was experimentally confirmed that high-order adjective pairs (see C1), which are general in color psychology, show extremely deep association and linearity. In order to show its deep relevance, examples of adjective pair terms that can be aligned from each of the sensitivity invariants are given below.

  It was also experimentally observed that the sensitivity axis gradually changed between the same types of sensitivity invariants with different quantum number differences. This is a description method very compatible with the two-faced nature of adjectives. In other words, adjectives, for example, have a subtle class of expressiveness such as “gorgeous”, “flashy”, and “noisy” for a major classification of “lively”, and their detailed classification ability is also a Kansei search. It becomes indispensable.

  The range of the effective quantum number difference is that when the non-diagonal component of the matrix element is smaller than the value that the diagonal component can take, that is, when it is relatively smaller than ± 1, there is no longer a meaningful sensitivity classification ability. It is appropriate to think of it not. That is, when the distribution range of the values of the sensitivity invariant, which is a spur of the matrix elements, is wide, alignment with high correspondence with sensitivity can be obtained, but when the distribution range becomes small, the correspondence relationship can no longer be seen. This has been confirmed experimentally.

[Reference C1] Color Society of Japan, Volume 1 “Color Science” (Asakura Shoten, 2004), Chapter 3 “Psychology of Color”, Section 3.2 “Measurement of Sensory, Perception and Cognition” ”, Table 3.4“ Adjective Pairs Commonly Used for Color Image Measurement ”and Section 3.4“ Color Recognition ”, Table 3.13“ Comparison of Factor Analysis for Color of Japanese and American Students ”
Specific examples are shown below.

  This is the arithmetic mean of the sensitivity invariants generated on each of the three color planes. From the sensitivity invariants, the sensitivity images of the adjective pair “static-dynamic” are aligned. “Static” images include distant landscape photographs that seem to stop for a moment. On the other hand, “dynamic” images include photos of a large number of people dancing at a festival and photos of the bustle of the city.

FIG. 8 shows an example of the color histogram shape of the HVC plane of an image located near both ends of the image distribution. The color histogram of a “static” image is collective with a shape like one band structure in which all of the H, V, and C planes are collected in one basis function state component. A “dynamic” image has many complex peak shapes on each of the H, V, and C planes, and has a histogram structure that is unlikely to be described with a simple waveform and is dispersive.

  From the sensitivity invariant, the sensitivity images of the adjective pair “tightened-open” are aligned. In the “tightened” image, images having many shadows are collected. On the other hand, a lot of images with a little brightness and spatial extent are gathered in the “open” image as a whole.

FIG. 9 shows an example of the color histogram shape of the HVC plane of an image located near both ends of the image distribution. The color histogram of the “tight” image is very similar in shape between VCs. On the other hand, the color histogram of the “open” image has completely different shapes on the V and C planes.

  From this sensitivity invariant, the sensitivity images of the adjective pair “excited-sedated” are aligned. This adjective pair is positioned in a conjugate relationship with the "tight-open" adjective pair derived from the symmetric product. “Excited” images include fancy gingko and maple trees and leaves, or a photo that captures the pink glow of a moment in the evening of a landscape photo and a moment when clouds appear. Many photographs that capture the movement of On the other hand, the “calm” image captures the moment when the movement is completely stopped and collects photographs with deeper colors.

<Model learning of sensitivity adjectives>
The value of the sensitivity invariant is uniquely determined from the observed amount of the distribution function of the image. It is necessary to determine a linear combination parameter for each adjective in order to obtain a correspondence with the adjective by pre-learning.

1) Least square method For one or more adjectives, an image having the impression is selected from a group of images for learning data. A square error function for measuring the reproducibility of the parameters is introduced, and each coupling parameter is determined in such a manner that each parameter is partially differentiated to find a minimum point, assuming that the linear coupling parameter is unknown. This is a coefficient fitting by the least square method.

2) Determination from positional relationship in image distribution gamut When multiple sensitivity invariants are generated from the distribution function of an image, which position is among the sensitivity invariants obtained for all images in the image database group. Is the nature of the sensibility itself produced by the image. That is, the boundary line of the distribution of the image group for each sensitivity invariant can be regarded as a gamut that the signal distribution of the natural image can take. The image located at the edge of the gamut may be considered to emit an extremely important signal for the invariant, and the image located in the middle is irrelevant to the invariant. It can be said that there is. Therefore, it is only necessary to set a numerical value in the range of [−1, 1] as the value of the linear combination parameter with the position where the image exists in the gamut as it is. However, normalization between all linear coupling parameters shall be performed last.

As described above, when a plurality of images are selected for learning with respect to one adjective, the coordinate position obtained by simply statistically averaging the selected image group within each sensibility invariant gamut is used as the adjective. The learning result of the linear combination parameter for. If a group of images selected for a given sensitivity invariant is scattered, the value of the parameter α _i approaches zero by statistical averaging, which means that the sensitivity invariant is irrelevant for the adjective. On the other hand, when the selected images are gathered in the same direction, even if statistical averaging is performed, a significant value remains in the parameter α _i , and the sensitivity invariant is important for the adjective. In this way, a sensibility invariant that works specially for an adjective can be derived very simply. The model learning method described above can be commonly used in all the embodiments described below.

[Second Embodiment]
(Linear sum of texture PDF Hilbert space representation and sensitivity invariant)
Next, generation of a sensitivity invariant according to the second embodiment will be described. In the second embodiment, the method for generating the sensitivity invariant in the first embodiment is changed to the following method.

<Processing for search target images>
1. Conversion to Munsell HVC color space As the hue plane, a plane that does not separate N is prepared. The hue plane of the N region behaves like random noise in the hue plane, but in the following edge extraction process, it plays a role of being detected as a feature that is different from other chromatic hues.

  The way of taking the one-dimensional coordinates in the hue circle may be to make red the origin of the Munsell hue circle around the start point and make the end point red magenta again, but more preferably as in the first embodiment. It is preferable to cut a point where the hue distribution frequency is minimum in each image and set a start point and an end point there. This is because the signal intensity fluctuates at both ends due to the division of the hue ring, and adverse effects due to overestimation of edge components when edges are extracted within the color plane are minimized.

となる。なお、LL成分は順次解像度の低い高周波サブバンドに分解されていくので、最終的に残るLL成分は最低解像度のみとなる。ウェーブレット変換としては、例えば以下のような5/3フィルタなどを用いる。 2. Creation of edge image 2-1. Multi-resolution transformation and edge extraction 1) Wavelet transformation The wavelet transformation is used to project a high-frequency edge component of each color plane of HVC by projecting to the frequency space expressed in multi-resolution. Here, it is assumed that the wavelet-decomposed high-frequency subbands LH, HL, and HH are used as they are as edge components. If you write this state schematically, when disassembling to resolution M stages

It becomes. Since the LL component is sequentially decomposed into high-frequency subbands with low resolution, the only remaining LL component is the lowest resolution. For example, the following 5/3 filter is used as the wavelet transform.

<Wavelet transform: Analysis / Decomposition process>
High-pass component: d [n] = x [2n + 1]-(x [2n + 2] + x [2n]) / 2
Low-pass component: s [n] = x [2n] + (d [n] + d [n-1]) / 4
The one-dimensional wavelet transform defined above is subjected to wavelet decomposition by performing two-dimensional separation filter processing independently in the horizontal and vertical directions. The coefficient s is collected on the L plane, and the coefficient d is collected on the H plane.

2) Laplacian pyramid As another method of multiresolution conversion, there is a method using Laplacian pyramid in addition to wavelet conversion. When creating a Laplacian pyramid, once create a reduced image (vertical / horizontal) * (1/2), restore it to the original size by bilinear scaling, and take the difference from the image before reduction. To obtain a high-frequency image (Laplacian component) of that resolution. Note that smoothing for preventing aliasing may be performed before a reduced image is created. If this is repeated one after another, a Laplacian pyramid with a series of high-frequency images is formed. As with the wavelet transform, only one low frequency image (Gaussian component) remains at the lowest resolution.
Document D1 describes that a histogram of high-frequency band signal values generated by multi-resolution conversion in this manner (called a probability density function, abbreviated as PDF) has a Gaussian distribution or a Laplace distribution. In general, PDF distribution can be approximated by symmetrical Generalized Gaussian.

The value of the multi-resolution conversion stage number M may be decomposed to a point where the PDF histogram of each band has a pixel number that is not rough. For example, it may be about 5 stages for Quad VGA size (1280x960) images, about 3 stages for QVGA size (320x240) images, and about 7 stages for 20 million pixel images.
FIG. 10 is a diagram showing a state of subband division by four-stage wavelet transform. For example, in the first-stage wavelet transform, high-pass component data and low-pass component data are first extracted from all rows in the horizontal direction with respect to real space image data. As a result, the data of the high-pass component and the low-pass component with half the number of images in the horizontal direction are extracted. For example, the high-pass component is stored on the right side of the memory area where the real-space image data was stored, and the low-pass component is stored on the left side.

  Next, high pass component data and low pass component data are extracted for all columns in the vertical direction with respect to the high pass component data stored on the right side of the memory area and the low pass component data stored on the left side. As a result, data of a high-pass component and a low-pass component are further extracted from each of the high-pass component on the right side of the memory area and the low-pass component on the left side. The high-pass component and the low-pass component are stored in the lower side and the upper side of the memory area where the respective data exist.

  As a result, the data extracted as the high-pass component in the vertical direction from the data extracted as the high-pass component in the horizontal direction is represented as HH, and the data extracted as the low-pass component in the vertical direction from the data extracted as the high-pass component in the horizontal direction Is represented as HL, data extracted as a high-pass component in the vertical direction from data extracted as a low-pass component in the horizontal direction is represented as LH, and is extracted as a low-pass component in the vertical direction from data extracted as the low-pass component in the horizontal direction. The data is denoted as LL. However, since the vertical direction and the horizontal direction are independent, it is equivalent even if the order of extraction is changed.

  Next, in the second-stage wavelet transform, the high-pass component and the low-pass component are similarly applied to the data LL extracted as the low-pass component in the vertical direction from the data extracted as the low-pass component in the horizontal direction in the first-stage wavelet transform. Perform extraction. If this is repeated up to four stages, the result is as shown in FIG.

[Reference D1] Michael J. Gormish, "Source coding with channel, distortion, and complexity constraints," Doctor thesis, Stanford Univ., March 1994, Chapter 5: "Quantization and Computation-Rate- Distortion."
2-2. Multi-resolution integration The high-frequency subbands extracted as described above represent information about edges, textures, and contrasts at each resolution scale. In order to handle this information in an integrated manner, multi-resolution inverse transform using only high-frequency subbands is performed to perform edge integration. That is, the low-frequency subband LLM with the lowest resolution is excluded and all of these values are set to zero, and then the remaining high-frequency subbands are sequentially subjected to inverse wavelet transform. When this state is schematically written, an integrated edge component having the same resolution as the input image is set as E, and the following equation is obtained.

  In this integration stage, information on edges, textures, and contrasts in different layers is transmitted to another layer in consideration of the spatial positional relationship. When the Laplacian pyramid is used, the Gaussian surface with the lowest resolution is set to zero, and the remaining Laplacian surfaces are sequentially integrated.

3. Create one-dimensional distribution function of integrated edge
Create a histogram (PDF) of the integrated edge image extracted from each color plane of HVC. The number of histogram bins should be set to about -128 to 128 for the H, V, and C with the origin in between. However, it is assumed that each color surface of the HVC is represented by a gradation of about 200 bins.

  Since PDF is a histogram of edge strength, it has a distribution with a peak at the origin having a frequency integration area of the same degree as positive and negative. In general, in the case of a memoryless source having no correlation between resolutions, a symmetric PDF distribution shape in each layer is integrated into a symmetric PDF distribution shape as it is. However, when there is a correlation between resolutions, the state of the correlation can be projected in the form of a PDF distribution shape.

In this way, the PDF of each high-frequency subband surface can be approximated by “generalized Gaussian: exp (-| x | ^α )”, but the spatial contrast correlation is reflected on the sequentially integrated edge surfaces. It changes to various shapes including asymmetry.

  It has been experimentally confirmed that such a characteristic shape of the PDF distribution of the integrated edge appears almost when the edge components of about three steps from the lowest resolution are integrated. Therefore, if it is desired to simplify the process, the PDF distribution shape in the middle of the integration may be evaluated without integrating the final actual resolution.

For convenience, if the histogram value is normalized by the number of pixels, a one-dimensional distribution function representing the probability density of the pixel value is obtained. The distribution function schematically created is expressed as follows. The Laplacian symbol Δ is used because the integrated edge image describes the secondary differential aspect of the pixel value of the original image.
f (△ H), f (△ V), f (△ C)
4). Hilbert space representation of distribution function The distribution function of the integrated edge image is series-expanded with a spherical Bessel function, and the shape can be evaluated by the expansion coefficient. In the second embodiment, right development and left development are performed separately, but in the third embodiment described later, left and right simultaneous development is performed. At that time, in the second embodiment, expansion by the root of the zero-order spherical Bessel function is performed. Fixing the outermost points of the development area and increasing the number of roots contained therein corresponds to generating a basis function having a high frequency component.

4-1. Variable transformation For the portion on the right side of the peak of the histogram, the distribution area on the horizontal axis is [a, b] (a <b), and the distribution area on the vertical axis is [fa, fb]. Variable conversion is performed in the interval where the horizontal axis is in [0,1] and the vertical axis is in [0,1]. Similarly, the left side of the histogram peak is set to [b, a] (b <a) on the horizontal axis and [fa, fb] on the vertical axis, and the same conversion is performed. Usually takes values of a to 0, fa to 0. For convenience only in this section, if the variable on the horizontal axis is converted from x to y and the variable on the vertical axis is converted from fx to fy, the conversion formula is as follows.
Variable transformation on the horizontal axis: y = | xa | / | ba |
Variable conversion on the vertical axis: fy = (fx-fa) / (fb-fa)
4-2. Series expansion by the root of the spherical Bessel function The distribution range of the horizontal axis subjected to the above-described variable transformation is represented by x regardless of ΔH, ΔV, and ΔC. The distribution function of each HVC color plane is expanded with a basis function based on the roots of N zeroth-order spherical Bessel functions. The meaning of the symbol a used here is different from the above, and means the outermost contour point that is the target of development of the distribution area.

The expansion coefficient c _n is obtained by the following equation using the orthogonality of the basis functions.

Here, α _nm means the value of the mth zero of the nth order function.

The root of the zeroth order function can be given analytically.
α _0m = πm, m = 1,2,3, ...
Therefore, the m-th basis function of the zeroth order function used for the series expansion has m zeros (roots) in the interval [0, a]. That is, the first zero point position of the 0th order spherical Bessel function with a set as the basis function of the lowest frequency in that order, the function is reduced toward the origin, and the second zero point is at the position of a When it comes, stop it and make it the second lowest frequency basis function in that order, and repeat this one after another to generate a high-frequency basis function in the interval [0, a] in the distribution range, forming a complete system Like that. This is exactly the same for the n-order function.

  Since the spherical Bessel function is a complete system in terms of root expansion, the original function can be completely reproduced by taking a sufficiently large value of N. When the number of bins in the histogram is about 128 on one side, the expansion order may be set to about N = 100.

5. Generation of Kansei Invariant Extract the shape of the distribution function of the texture PDF on each face of the HVC by the synthesis system of two basis functions. That is, the sensibility that appears due to the structure of each distribution function of the HVC integrated edge image is extracted as a feature amount.

  Since one distribution function obtained from edge components in the same color plane is expanded to the right and left using the same basis function, the method of creating a quadratic invariant is more than that of the first embodiment. The number of types increases more than twice. The following shows an example of how to create an invariant when the quantum difference is 0 and 1 in the case of a symmetric product, and how to make an invariant when the difference is 1 in the case of an antisymmetric product. Products can be defined with quantum difference up to m = 0,1, ..., N / 2, and antisymmetric products with quantum difference up to m = 1,2, ..., N / 2-1.

Considering the right and left sections of the distribution function as different quadrants, a composite system can be created in the same way as in the first embodiment from the combination of two basis functions in the same quadrant on the HVC3 plane, and between different quadrants. A composite system can also be created from a combination of two basis functions. Since the former handles only the radial direction, the symbol “r” is given. The latter gives the symbol “z” in the sense that it crosses zero. Let the expansion coefficient on the right be c _k ^{(α +)} and the expansion coefficient on the left be c _k ^(α−) .
Here, with one exception, it is standardized and defined so that all invariant values fall within [-1,1]. Further, when the sum range exceeds the range of k = 1,..., N, it is treated that k = 1 is connected in a ring shape after k = N. The color plane is (α), (β) = H, V, C.

Evaluation of basis functions in the same color plane by the synthesis system 1) Combinations in the same quadrant

2) Combination of different quadrants

Evaluation of basis functions between different color planes by synthesis system 1) Combinations in the same quadrant

2) Combination of different quadrants

  The values that can be taken by the invariants and their properties are the same as those described in the first embodiment, and are therefore omitted here.

6). Creation of adjective discrimination index 6-1. Linear combination of Kansei invariants Creates a new index Q _i that linearly combines Kansei invariants using the additive property of Kansei invariants as an index for searching for Kansei adjectives (i) . The adjectives that can be represented by the index Q _i are not only one adjective but also a pair of adjectives that have adjectives of the opposite nature.
Q _i = β ₁ G ₁ + β ₂ G ₂ + ...
Here, the value of the linear coupling parameter β _i is normalized so that Q _i becomes an index in the range of [−1, 1] again.

6-2. Parameter setting for search target adjectives Linear combination parameters corresponding to predetermined adjectives are learned in advance, and the model parameters are set.

7). Image Classification Processing Based on the adjective discrimination index, images are classified as in the first embodiment. In the following, a specific example is given to show what kind of sensibility is actually given. Similarly to the first embodiment, in the case of texture, from each sensitivity invariant, it was possible to obtain an image alignment having a very deep relationship and linearity with high-order adjective pairs used in color psychology.

  An overview of the nature of sensibility invariants shows that the invariant that contains the asymmetry of the distribution function is considerably stronger in emotional elements by combining basis functions in different quadrants. On the other hand, when the basis functions in the same quadrant are combined, separation such as multiplicity and unity of the object structure of the image is possible, but a fairly neutral aspect is also strong. In the following, examples are given mainly by combining different quadrants while assigning adjectives.

  The situation is slightly different from that of the color distribution function of the first embodiment. In the case of the edge distribution function, a large value remains even when the element of the off-diagonal component is away from the diagonal region, and the former is Compared to the short-range correlation whose strength decreases rapidly when leaving the angular range, it has a long-range correlation property. This can be interpreted as indicating that the number of emotions induced by the texture distribution structure is greater than the number of emotions induced by the color distribution structure.

の相加平均をとった場合に、展開の次数がN=100で量子数差mをm=0,8,50とずらしたとき、整列させた画像群分布の両極端にはどのような分布関数の形状が該当するかを示した例である。上段は一方の端に位置する画像のV面のエッジ画像の分布関数の様子を、下段はもう一方の端に位置する画像の同様の様子を示している。 In FIG.

What is the distribution function at the extremes of the aligned image group distribution when the expansion order is N = 100 and the quantum number difference m is shifted to m = 0,8,50 when the arithmetic mean of is taken? It is an example which shows whether the shape of corresponds. The upper row shows the distribution function of the edge image of the V plane of the image located at one end, and the lower row shows the same state of the image located at the other end.

As the tendency of the image group corresponding to FIG. 11, images with large elements in the image at the upper stage of m = 0 gather together to give a “rich” impression, and images at the lower stage give a “mysterious” impression. The tendency to gather was recognized. Also, from m = 8, there are many composite images in which two subjects overlap in the front and back in the upper part, and there are two elements of the background and the main subject, and the leaves and grass are reflected on the entire bottom part. Many integrated images with a fine texture structure are collected.

From the sensitivity invariant, the sensitivity images are arranged such that the two adjective pairs “lively-lonely” and “gorgeous-clean” are combined. “Buzzy, gorgeous” images are a collection of images of various scenes that include many large and small structures. On the other hand, the “lonely, neat” image contains many photographs that contain dark-colored structures and dark shaded scenes with an area specific gravity that has some visual influence.

  From the sensitivity invariant, the sensitivity images of the adjective pair “smudged-mysterious” are aligned. Many images with trees and shades are gathered in the “smeared” image. On the other hand, the "mysterious" image is accompanied by a picture of a large area, a section of landscape, a dyed pink, orange, yellow, etc. at dusk, or a series of strange clouds. Many images like steam and haze are collected.

FIG. 12 shows a typical edge image and distribution function of images selected at both ends. Explaining from the shape of the distribution function, as shown on the left side of FIG. 12, the V plane and the C plane both have a lot of edges and textures, giving a familiar impression. On the other hand, as shown in the right side of FIG. 12, the distribution function of the V-plane edge image shows a strong value with a slight frequency, that is, if there is a small base in the distribution function, the original picture is divided into sections such as mountain ridges. It is considered that a ghostly impression is given by the fact that the change of the saturation plane is woven with a thick contrast width larger than the contrast change of the luminance component.

  Although it is difficult to apply an adjective pair from this sensitivity invariant, it is clearly separated into a pair of images of “summer and day scenery” and “autumn, spring and evening scenery”. The image of “summer and daytime scenery” has a lot of green and blue areas and a strong contrast. “Autumn / Spring / Evening scenery” images contain warm colors, and there are many images with little change.

[Third embodiment]
(Linear sum of texture PDF Hilbert space representation and sensitivity invariant)
Next, a third embodiment will be described. In the second embodiment, the histogram of the integrated edge image is developed separately on the right side and the left side. However, it is easier to discuss the symmetry when the left and right sides are developed simultaneously, so only the points to be changed are described below. It is considered that the sensitivity invariant obtained therefrom is more easily correlated with sensitivity than in the second embodiment and has a higher degree of alignment.

4). Hilbert Space Representation of Distribution Function A distribution function is represented in Hilbert space by a double series of root expansion by a zero-order spherical Bessel function that is an even function and a root expansion by a first-order spherical Bessel function that is an odd function. Then, all the symmetric components of the distribution function shape are collected in the expansion coefficient of the zero-order function, and all the asymmetric components are collected in the expansion coefficient of the linear function. Further, the double series may be expanded to an infinite order by further development.

4-1. Variable transformation Let r be the distance from the peak position p of the histogram to the extreme end of the distribution area that extends to the farthest right or left. The distribution range on the horizontal axis is [-r + p, r + p]. Variable transformation is performed on the horizontal axis in the interval [-1,1].
Variable transformation on the horizontal axis: y = (xp) / r
The vertical axis is the same as in the second embodiment.

4-2. Double series expansion of roots and orders of spherical Bessel functions
The distribution function of each HVC color plane is expanded with basis functions based on the roots of N zeroth-order and first-order spherical Bessel functions. The meaning of the symbol a used here means the outermost contour point that is the target of development of the distribution area, as in the second embodiment.

The expansion coefficient c _nk is obtained by the following equation using the orthogonality of the basis functions.

Here, when guiding the equation for calculating expansion coefficients, the product of the zero-order and first-order function since the internal variable is an odd function be how scaled, and further consideration of the even function of [rho ^2, left and right We used the property that the integration value becomes zero when integrating over a symmetrical interval. That is, the relational expression of the orthogonality regarding the order is not limited to the case where the weighting function is 1 as long as it is defined between the even-order function and the odd-order function, and can be applied even if the weight of an arbitrary even function is added. However, the constant factor of normalization will vary but, (described in the portion of the outline description) that already have been given by the equation according to the roots in the case of [rho ^2.

5. Generating Kansei Invariants If the double expansion coefficient c _nk is treated as one vector in the same way as the single expansion coefficient, the Kansei invariant G _{i in the} same form as the color Kansei invariant F _{i of} the first embodiment is used. Can be built. Therefore, it is only necessary to replace F with G. However, the range taken by the sum is doubled and changed to k = 1, 2,..., N, N + 1,. Image classification can be performed in the same manner as in the first and second embodiments using the sensitivity invariant thus obtained.

[Fourth Embodiment]
(Linear sum of color and texture sensitivity invariants)
Next, a fourth embodiment will be described. The sensitivity invariant F _i derived from the one-dimensional distribution function related to the color signal distribution of the image shown in the first embodiment, and the one-dimensional value related to the color signal distribution of the edge image shown in the second and third embodiments. Kansei invariant G _i led from distribution function, from having There additive nature in exactly the physical quantity of the same dimension, both by the linear combination can be handled in common. This is not just a one-dimensional distribution function, but the two-dimensional distribution function of the image color signal distribution itself and the higher-order sensitivity elements of the image entangled with the composition elements are completely the same if the sensitivity invariants are constructed in the same way. It becomes possible to explain the property that the superordinate concept remains strongest in the hierarchical structure of adjectives by using a linear model of sensitivity.

<Processing for search target images>
1. Conversion to Munsell HVC color space In this fourth embodiment, two hue planes are prepared: a plane with N separation and a plane without. Subsequent processing is the same as in the first to third embodiments.

5. Creating an adjective discrimination index
Q _i = α ₁ G ₁ + α ₂ G ₂ + ...
+ β ₁ G ₁ + β ₂ G ₂ + ...
<Sensitive adjective model learning>
The linear coupling parameter is determined by the same method as in the first embodiment.

Here, we will summarize the previous explanations and give a general overview to guide new concepts. After that, fifth to seventh embodiments for realizing them will be described.
[1] Image of Kansei Formulation <Basic Concept>
Consider the following as a picture of Kansei formulation. Each image emits an energy element common to the sensibility. The brain perceives it.

<Shape characteristics of distribution function>
FIG. 13 is a diagram illustrating a distribution function of color and texture. As described so far, the sensitivity of the image and the shape of the distribution function are very related, and the shape similar to the ambiguity tends to give the same sensitivity. Physical techniques have been introduced to quantify the shape recognition of this distribution function.

<Quantum mechanical method>
The contents explained so far are to try to describe by quantum mechanics. That is, it can be summarized as follows.
1) The distribution function f is projected onto the Hilbert space and the momentum p is displayed.
2) constructing an additive energy E _n of the quadratic form with group theory.
The method adopted here actually introduces the concept of quantum mechanics, and describes the state of a multi-system with a large number of pixels and a large number of gradations as a statistical mechanics through a distribution function f. Also included. This will be described in detail below.

<Statistical physics>
The meaning of the above description method is conversion from a physical quantity having a micro property of image signal value S (x, y) to a physical quantity having a macro property of sensitivity. In other words, by taking a statistical average of micro quantities, only mechanical invariants such as momentum, angular momentum, and energy remain as effective components. Among them, by taking a statistical average of a group of images, an image in which only energy is effective is obtained. Plays a role in characterizing the system. By the action of the statistical group, the image system is aggregated in the form of energy band structure. By reducing the information volume of the image system in this way, conversion from the micro property to the macro property is performed. Statistical mechanical description serves as a bridge to describe its statistical properties.
The state of information reduction is shown below. The composition system is incomplete, but the prediction is included.

Image signal S (x, y) Sub system information f (p, q) Energy information ~ 10 ^ 23 colors: (256) ^ 3 ~ 10 ^ 7 ~ 2000
Texture: (± 256) ^ 3 ~ 10 ^ 8 ~ 2000
Composition: Number of pixels ~ 10 ^ 7 (~ 2000)
Here, the variables p and q are symbols that represent the independent variables p1, p2, ..., pi, ..., q1, q2, ..., qi, ... Used as

<Description of quantum statistics by density matrix>
In describing the dynamics, it is necessary to consider what corresponds to the position coordinate q and the momentum p with respect to the image signal S (x, y). Before that, the relationship between the image system and quantum statistics will be explained.
Sensitivity deals with unknown Hamiltonian systems and dynamical systems made up of statistical groups. As known in statistical mechanics, there cannot be a wave function that describes the entire macroscopic system (see reference E1). The image system also describes a macroscopic system.

A quantum mechanical description based on an incomplete set of data about a system is done using a density matrix. The expected value of any physical quantity can be calculated using the density matrix. The density matrix for coordinates is
ρ (q, q ') = Σ _{m, n} w _{m, n} ψ _m * (q') ψ _n (q)
It is represented by If a subsystem related to a certain aspect of an image is in a state completely described by a wave function ψ, the wave function ψ can be expanded by a function ψ _n (q) that forms a complete system.
ψ = Σ _n c _n ψ _n
Substituting this into the density matrix expression gives the density matrix in the energy display.
c _m * c _n → w _mn
Can guide you.

The diagonal component of the energy display density matrix represents a steady state. One off-diagonal component represents an unsteady state. The image captures the dynamism of a moment. It is not always steady state. Therefore, the energy levels derived from the density matrix must also take into account off-diagonal components. This means that the description of the off-diagonal component depends on how to select the wave function that describes the density matrix. Choosing a wave function system in which the steady state of the diagonal component is described as compactly as possible is an appropriate choice for describing sensitivity. In statistical physics, the steady state is usually considered, so the probability of being in the nth steady state is written as w _n = w _nn .

  The statistical average also calculates the expected value in steady state. However, the sensitivity of the image is not limited to the steady state, and the non-steady state needs to be considered. Therefore, it is necessary to calculate the expected value in consideration of both the steady state and the unsteady state.

When dealing with the ambiguity of sensibility that several image groups give the same sensibility, in the description of sensibility, the steady state is defined as one energy level state (E ₀ ) by taking the diagonal sum of all steady states. Similarly, the unsteady state is described by the energy level (E _n ) of how far away from the steady state to the unsteady state by taking the diagonal sum of the non-diagonal components. Take the way. The statistical distribution of the excitation probability to the energy level is represented by w _n = w (E _n ). From a physical point of view, it can be said that as the value of n increases, it describes a more dynamic state of motion in the image system and an image state in a changing process with a large energy transition width. This argument is applied as it is in the process of building the energy matrix.
If energy is defined and their state density distribution is expressed by ρ (E _n ), the actual energy probability distribution is ρ (E _n ) along with the excitation probability w (E _n ) to the energy level. _n ) w (E _n ).

[Literature E] Landau, Lifschitz Theoretical Physics, Vol. 5 "Statistical Physics Part 1" (3rd edition, 1976), Chapter 1 "Basic Principles of Statistics", Chapter 2 "Thermodynamic quantities" , And Chapter 3, “Gibbs Distribution”
[Literature E1] Landau, Lifschitz Theoretical Physics, Vol. 5, "Statistical Physics Part 1" (3rd edition, 1976), Chapter 1, "Basic Principles of Statistics", Section 5, "Statistics Matrix"
[2] Relationship between distribution function and Kansei <Subsystem and distribution function>
In describing sensitivity, finding an statistical distribution for a subsystem is an important issue. Coordinates and momentum dynamic variables q and p in sensitivity are not necessarily the only ones defined by dynamics. As a sensibility sub-system, we project a certain aspect of the image and describe the nature of the image as a system that captures the highly independent aspect.

  Three major axes of color, texture, and composition can be considered as a major classification of that aspect. However, since the projection is based on the same image, it is not completely independent. However, in defining the image information, if the distribution functions of these three aspects are defined, the statistical properties are almost described. You will have enough information.

  Here, we will keep the simple description level of dynamical systems and devise a description method that is not completely independent when a more advanced description is required. In a sense, this is close to the relationship between a non-relativistic description of physics and a relativistic description. In other words, in the non-relativistic description, the position coordinate system and the spin coordinate system are described as completely independent systems, but in the relativistic description, the distinction is impossible, and the coordinate system in which the position coordinate and the spin coordinate are mixed. It is because it is similar to the situation that it is necessary to shift to the description by Spinor. After describing each of these subsystems, the overall sensitivity of the image system is described by integrating them.

<Description of distribution function and dynamic variables of subsystem>
The image variables corresponding to the position coordinates q and the momentum p, which are mechanical variables, are defined differently for each sub-system. The color represents the distribution of signal values with an interval width of "zero to positive finite value" and describes the distribution of the original signal itself. Therefore, a function that satisfies the linear differential equation suitable for the description is a hypergeometric function. Any function included. The texture represents the distribution of the signal value with the interval width from “negative finite value to positive finite value” from which the edge component is extracted, and describes the distribution of another aspect from which the original signal information is dropped. The function satisfying the linear differential equation suitable for description is any function included in the confluent hypergeometric function. Since the composition represents a two-dimensional distribution of signal values with an interval width of “zero to positive finite value” and describes the distribution of the original signal itself, the function satisfying the linear differential equation suitable for the description is Any function included in the hypergeometric function.

  It can be considered that the system described by the hypergeometric function and the confluent hypergeometric function captures subsystems with different aspects. That is, the hypergeometric function represented by a differential equation having three singular points is suitable for representing the distribution of the signal value of the image itself, and the image signal that comes to the argument of the distribution function is zero as the pixel value. It is suitable for describing the property that an image signal composed of the above values can be uniformly distributed. On the other hand, the confluent hypergeometric function expressed by a differential equation having two singularities in total by merging the two singularities into one has the aspect of reducing the amount of information about the signal value of the image by one. Suitable for description, as an image signal that comes to the argument of the distribution function, an image signal that spans positive and negative values, such as an edge signal value from which one piece of information has been dropped by differentiation, is localized near zero. Suitable for describing properties to be

<Image of variable separation of mechanical variables>
Describe the side of the image sub-system by drawing a picture corresponding to the coordinate system with variable separation in the description of physical particles. The hypergeometric function suitable for describing the one-dimensional color distribution function is the Chebyshev function. Since this function forms a complete system with a single series of even and odd functions, the even function is angular momentum 0 and the odd function is angular. If it is regarded as momentum 1, it can be considered that the spin coordinates of spin 1, which is the same Bose particle as light, are described. Therefore, it can be considered that the one-dimensional distribution function of color describes a wave function of a spin system among the wave functions.

  A confluent hypergeometric function suitable for describing a one-dimensional distribution function of a texture is a spherical Bessel function. Of the radial wave functions normally defined in the positive region, the spherical Bessel function is the only function that can be expanded and defined in the negative region of the Bessel function. Therefore, it can be understood that the one-dimensional distribution function of the texture describes the radial wave function among the wave functions. Since the radial wavefunction can be described by a double series of two series related to the order and root of the even and odd functions, the angular momentum 0, 1, 2, 3, ... They can be mapped to what are called s, p, d, f, ... orbitals in atomic orbitals.

  The hypergeometric function suitable for describing the two-dimensional distribution function of the composition is considered to be a Legendre power function or a spherical harmonic function that is a product of a Fourier function with the Legendre function. Therefore, it can be considered that the two-dimensional distribution function of the composition describes a wave function corresponding to the two-dimensional system coordinates of the zenith angle and the azimuth in the spherical coordinate display.

<Observation data set and distribution function>
Certain aspects of the image information are described by the distribution function f (p, q). The relationship between the data observed as image information and the distribution function is positioned as follows. That is, color: a distribution function in which information on pixel positions is lost by taking a histogram, texture: a distribution function in which information on pixel positions is lost by taking a histogram on the side of an edge, and composition: the number of pixels by taking a region average, that is, pixels Position like a distribution function with reduced information about position. Together, these three distribution functions accurately reflect the statistical nature of the image S (x, y). There is a low probability that another image with a different impression will be represented in the same value. Here, what is described as a pixel position cannot be generally said to correspond to the position coordinate q of the dynamic variables p and q. It will be described below that two spaces can be defined for each color and texture.

<Coordinate space and momentum space>
(A) Side surface of low-order space
(A-1) Color Munsell color space can be regarded as a momentum space representing the intensity distribution of colors.
Color histogram = f ₁ (p)
(A-2) Texture The color space of each color plane of the HVC of the edge image obtained by integrating the multi-resolution edges can be regarded as a momentum space representing the edge intensity distribution.
Texture PDF = f ₃ (p)
(B) Side of higher space
(B-1) Color When the color histogram is subjected to Chebyshev transformation, the original color histogram side can be regarded as the position coordinate q, and the Chebyshev expansion coefficient side can be regarded as the momentum p. The distribution function projected onto the higher-order momentum space needs to represent the probability density and does not take a negative value, so the expansion coefficient side takes a power spectrum obtained by squaring the coefficient.
Color histogram = f ₂ (q) = f ₁ (p)
Chebyshev spectrum = f ₂ (p)
(B-2) Texture When texture PDF is transformed into a spherical Bessel, the original PDF side can be regarded as position coordinate q, and the spherical Bessel expansion coefficient side can be regarded as momentum p. Similarly, a power spectrum is taken as a distribution function of the higher-order momentum space.
Texture PDF = f ₄ (q) = f ₃ (p)
Spherical Bessel spectrum = f ₄ (p)
FIG. 14 is a conceptual diagram for easy understanding of the relationship between them. These distribution functions are used to calculate the entropy of each subsystem defined later. When creating a quadratic form in energy construction, the description of kinetic energy in the usual Cartesian coordinates (x, y, z) ((p _x ^ 2 + p _y ^ 2 + p _z ^ 2) / (2m) In contrast to the correlation within the corresponding color plane, it is meaningful to create and evaluate correlation elements between color planes, because Munsell HVC is not a completely independent component. Is working. However, since the Munsell HVC color space is designed as a uniform color space having an equal rate with respect to the psychological quantity, there is no need for a coefficient between each quadratic term.

<Two meanings regarding statistical average>
The statistical average for deriving the statistical properties of the image system has two meanings. That is, one point is that a huge amount of information on color, texture, and composition is reduced to a macroscopic physical quantity by a statistical average. Another point is that a definite value is obtained by taking an average of image groups in the sense that a large number of model image groups have a common sensitivity.

  The former includes a quantum mechanical average operation and a statistical average operation at the same time, accompanied by a conversion process from a microscopic physical quantity to a macroscopic physical quantity. The latter is mainly regarded as including a statistical averaging operation for determining a statistical property by determining an amount corresponding to an average value of fluctuations of a macroscopic physical quantity. Therefore, the former makes a signal that makes a macroscopic action on the brain through the visual system, and the latter plays a role of knowing the average element of the distribution of actions in the brain of the sensation received by a person.

[3] Description of sensibility and Gibbs distribution <Fluctuation of energy created by images>
The method of finding the statistical properties of the common sensibility of a large number of images corresponds to the Gibbs distribution (canonical distribution) of statistical physics. In other words, it is a system description method that takes into account energy fluctuations. On the other hand, the method for describing the statistical properties of a single image corresponds to the microcanonical distribution. This is a system description method that ignores fluctuations in energy.

  The momentum and angular momentum of a closed system (single image) are related to uniform translational motion and uniform rotational motion of the entire system. It does not describe the statistical properties of the system. However, it is meaningful in the sense to distinguish it from the uniform translational motion and rotational motion of other images. That is, if a coordinate system in which the translational and rotational motions of a closed system of a single image are stationary is taken, only the energy determines the statistical properties of the single image, but the coordinate system common to other images Since the momentum and the angular momentum are different from those of the image group, they are recognized as different images. However, when describing the macroscopic properties common to many images, that is, sensibility, the momentum and angular momentum are information that disappears to zero or a constant value when the statistical average of the image group increases.

<Description of sensitivity ambiguity and phase space>
FIG. 15 is a pictorial diagram showing the sensitivity group in the phase space. Sensitivity description is also a question of how to describe ambiguity quantitatively. As a picture of the statistical physics description method, the locus of the phase space where the distribution function satisfies a certain condition describes the state distribution that evokes the same sensitivity. The constant condition represents a trajectory on the phase space that satisfies a condition in which energy, which is a mechanical invariant, is constrained to a constant value. The ambiguity of sensibility is described as a collection in which the distribution function can take various states as an image, but they are all on the locus of a blurred range in the same phase space. That is, even if the energy elements derived from the distribution function are not the same, if the sum of the energy elements satisfies the same condition, the same sensitivity is evoked. Moreover, since it has energy fluctuations, it is considered that trajectories having a certain width in the phase space correspond to the same sensitivity group. In FIG. 15, the state is drawn so as to be widened with a thick line.

  In general, the logarithm of a distribution function is an integral of motion, which is a mechanical invariant, and is described by an integral of additive motion. In other words, the distribution function of a composite system, which represents the probability that the distribution function of one subsystem and the distribution function of another subsystem simultaneously take that state, is represented by the product of the distribution functions of the subsystem, and its logarithm is an additive property. Because it has. Such additive integration of motion corresponds only to energy, momentum and angular momentum, as is known in mechanics. Accordingly, the logarithm of the distribution function can be described by a constant term α and a linear combination of the constant coefficients β, γ, and δ applied to each of them (see document E2 = document A3).

[Literature E2] Landau, Lifshitz Theory of Physics, Vol. 5, "Statistical Physics Part 1" (3rd edition, 1976), Chapter 1, "Basic Principles of Statistics", Section 4, "Role of Energy"
<Phase space orbit of a small number of models>
When the population image of the image group that evokes the same sensitivity is statistically sufficiently large, the constraint condition is only energy. This comes from the fact that only the energy remains as an integral of the additive motion in the description of the statistical properties (see document E2 = document A3).

  However, in the case of a small number of them, the corresponding impression of sensibility requires a stronger constraint that matches the fine distribution of the few models. The momentum and angular momentum, which are other mechanical invariants, play a role as additional constraints for that purpose, and in addition to energy, the trajectories on the phase space satisfying the constant momentum and angular momentum respectively. The distribution range exerts a sensibility effect common to a small number of model images. As a limit of the minority model, a similar image search of a single image is applicable.

<Momentum display and uncertainty principle>
FIG. 16 is a diagram for explaining a situation where the relationship between the position and the momentum satisfies the uncertainty principle in the projection onto the higher-order space regarding the color and texture of the image system. That is, as in the case of the texture PDF illustrated in FIG. 16, when a sharp peak is raised with respect to the position and the accuracy of position coordinates is high, any frequency can be obtained by trying to realize the waveform in a momentum space obtained by spherical Bessel-Fourier transform. Since it cannot be achieved without overlapping a large number of momentums, the momentum becomes uncertain and has a very wide distribution. On the other hand, when the Chebyshev Fourier spectrum, which is a momentum space, is expressed in a highly concentrated manner as in the case of the color histogram illustrated in FIG. 16, if it is expressed in real space, a wave of a certain frequency is generated in this case. Represents a distribution of low-frequency waves spread throughout the entire space. Therefore, the position becomes uncertain and the spread width is large. In other words, if frequency analysis is also performed in an image system, it can be rephrased as a relationship that describes the fact that the frequency components contained therein cannot be specified unless a waveform with a sufficient interval width is seen in real space. .

  Through such an extended Fourier transform concept between position and momentum, the uncertainty relation that the wave packet at the position and the wave packet at the momentum cannot be realized simultaneously beyond a certain width is discussed in quantum mechanics. This is nothing but the principle of uncertainty. An explanation of this principle in quantum mechanics is given, for example, in document F1.

  In the image system, the value of the plank constant h is related to the concept of a histogram bin that defines the gradation width with respect to quantization in the gradation direction and one pixel that defines the pixel interval with respect to quantization in the spatial direction. . Therefore, when the number of gradations of an image system is reduced or a reduced image is generated, the Planck constant value needs to be changed according to the quantization method of the system, and changes depending on the situation to be considered. This is different from the situation where the entire system can be treated as a constant regardless of the subsystem, such as physical constants. However, the Planck's constant in the image system can be treated as a fixed value in a sub-system in which a certain width is projected and the quantization width is defined.

[Reference F1] Landau and Lifshitz, Course of Theoretical Physics, Volume 3 "Quantum Mechanics (Non-Relativistic Theory)," (Third revised edition, 1977), Chapter 2 "Energy and Momentum," Section 16 "Uncertainty relations."
<Phase space and quantum statistics>
The trajectory in the phase space, which is a basic concept of statistical physics, allows the introduction of the concept of the number of states as a quantum mechanical correspondence. Image sensitivity can be quantified by the expressed density of states.

  When the number of quantum states divided by the Planck constant divided by the number of uncertainties in both momentum and coordinates can be defined on the phase space in terms of quantum theory, the additive property represented by the logarithm as its macroscopic property The concept of entropy, which is a complex physical quantity, inevitably appears statistically. Through this concept of entropy, the determination of the density density of the energy spectrum of a macroscopic system is based on statistical physics, and the number of particles (here, the number of pixels and the number of gradations in an image system) It can be explained that the level interval decreases exponentially as the response increases, resulting in a continuous energy band structure (see document E3).

As a result of actually calculating the continuous energy level E _n by experiment, there is an image alignment for energy elements that show almost similar properties between adjacent energy levels. In the distribution of energy elements, the change of order over long distances has begun, and when looking at the state of this property in order from end to end, the high energy level is completely different from the low energy level. Image alignment with different properties is obtained. This represents the nature of the energy band structure itself in which the energy level density is increased and the state between levels cannot be distinguished. Note that the Boltzmann constant connected by the relationship between the entropy and the Planck constant can also be treated as a fixed value in a partial system on which a certain side surface is projected, but in general, the definition differs between different subsystems.

[Literature E3] Landau, Lifschitz Theoretical Physics, Vol. 5, "Statistical Physics Part 1" (3rd edition, 1976), Chapter 1, "Basic Principles of Statistics", Section 7, "Entropy"
<Band structure of statistical group and energy level>
This results in an unusually dense level distribution in the spectrum of the energy eigenvalue of a macroscopic object (here, an image) (see reference E1). Therefore, information is gathered in the form of an energy band structure in the image system.

Here, the similarity between the theoretical structure of physical properties and sensitivity will be described.
Differences in properties between pure iron and iron alloys are expressed as differences in the energy band structure of electrons (see references G1, G2, and G3). Alternatively, how the differences in properties of ferromagnetic iron, nickel, and cobalt are born can be described by elucidating the electronic structure. The same applies to the difference between paramagnetic metal and ferromagnetic metal, or the difference in properties of paramagnetic metal such as aluminum and copper. The kind of substance is composed of 118 kinds of elements, their compounds, alloys, etc. as a simple substance, and exists in the order of thousands to 10,000. For example, if the number of combinations of 2 to 3 elements out of 118 elements is calculated, this number is obtained. The number of elements and the basis of their properties are defined by the radial wave functions s, p, d, f orbits that can exist as atomic orbitals and the degenerate numbers of those states.

  Similarly, there are typically 473 sensual adjectives that represent Japanese color emotions, which cannot be distinguished by those words. It is thought that it exists in the order of 10,000. The meanings of these representative 473 words are not always clearly different, and there are many expressions that express subtle differences. For example, there are different expressions for “lively” such as “lively”, “gorgeous”, and “brilliant”. On the other hand, there are adjective groups that are quite different from the viewpoint of the broad classification of “fresh” and “lively”. In this way, there are two types of adjectives: major classification and fine classification.

Such dihedral properties also exist in the properties of materials. For example, in the periodic table of elements, the vertical wave function is higher in the vertical direction, and the radial wave function is the s orbit and p orbit. Light electron systems packed in sequence, the intermediate zone is a transition metal system packed with d orbitals, the lower is a heavy electron system packed with f orbitals, and the horizontal axis shows one degenerate number of electrons in those orbitals. Different things are lined up. In terms of material properties, elements present in the same row in the vertical direction exhibit chemically very similar properties, and exhibit similar properties when present adjacent in the horizontal direction. These are statistically energy band structures when substances are aggregated as a large population that spans the Avogadro number (up to 10 ^ 23) as solids. Differences can also be expressed.
Therefore, it is an optimal expression method that the sensitivity is described by such an energy band model. In addition, when expressing the two-sidedness of adjectives in the phase space, the sensibilities that are different in the large classification are different in the state distribution between those that are greatly different in the trajectory, and the sensitivities that are different only in the fine classification have a considerable trajectory It is thought that it expresses the difference in the state distribution between close ones.

[Reference G1] Masako Akai, Hisazumi Akai and Junjiro Kanamori, "Electronic Structure of Impurities in Ferromagnetic Iron.I.s, p Valence Impurities," Journal of Physical Society of Japan, Vol.54, No.11, November, 1985, pp.4246-4256.
[Reference G2] Masako Akai, Hisazumi Akai and Junjiro Kanamori, "Electronic Structure of Impurities in Ferromagnetic Iron. II. 3d and 4d Impurities," Journal of Physical Society of Japan, Vol.54, No.11, November, 1985, pp .4257-4264.
[Publication G3] Masako Akai, Hisazumi Akai and Junjiro Kanamori, "Electronic Structure of Impurities in Ferromagnetic Iron. III. Light Interstitials," Journal of Physical Society of Japan, Vol.56, No.3, November, 1987, pp.1064 -1077.
[4] Description of macroscopic physical quantity <Image level imaging model>
Energy can be defined based on the definition of momentum and position in each subspace defined by projection onto the subsystem. Also, angular momentum, which is another mechanical invariant, can be defined. As energy building methods, two methods are introduced for each sensitivity principal axis: projection into a lower-order subspace and projection into a higher-order subspace. In the case of projection onto a lower-order subspace, a method is proposed in which a model Hamiltonian of energy is proposed. In the case of projection onto one higher-order subspace, an energy matrix is constructed and an energy eigenvalue from a steady state to an unsteady state is defined.

  The energy levels in the lower-order subspaces take the field energy created by the mean field approximation of statistical physics as discrete energy levels. That is, construct a discrete scalar invariant. The energy level in the higher-order subspace is considered as a continuous energy level in the process of gradually separating from the stationary state of the diagonal component of the energy matrix to the unsteady state of the non-diagonal component. That is, construct a continuous vector invariant.

<Low-order invariants of color>
The Munsell H, V, and C values themselves represent momentum.
The model Hamiltonian is constructed as follows.
H = (H + V + C) ^ 2
This represents the kinetic energy representing the intensity of the color value. Or it represents the energy of the field.
equation
Hlψ _n > = E _n lψ _n >
The other hand, finds the energy eigenvalues E _n.
In calculating the energy eigenvalue, the mean field approximation of statistical physics is used.
Each momentum is divided into an average term and a fluctuation term, and the fluctuation term is also described by an average fluctuation width. That is, the standard deviation is taken as the fluctuation term, and the dimension of the momentum p is the same as the average value.
√ <(H- <H>) ^ 2> 〜 = σ _H
√ <(V- <V>) ^ 2> 〜 = σ _V
√ <(C- <C>) ^ 2> 〜 = σ _C
The average value and standard deviation value of these momentums are obtained from the distribution function f (p) representing the momentum distribution. That is, it is obtained from the color histograms f (H), f (V), f (C) of H, V, C.
H = (H + V + C) ^ 2
= {(<H> + (H- <H>) + (<V> + (V- <V>) ++ C> + (C- <C>)} ^ 2
~ = {(<H> + σ _H ) + (<V> + σ _V ) + (<C> + σ _C )} ^ 2
= (<H><H> + <V><V> + <C><C>)
+2 (<H><V> + <V><C> + <C><H>)
+2 (<H> σ _H + <V> σ _V + <C> σ _C )
+2 (<H> σ _V + <V> σ _C + <C> σ _H )
+2 (σ _H <V> + σ _V <C> + σ _C <H>)
+ (σ _H σ _H + σ _V σ _V + σ _C σ _C )
+2 (σ _H σ _V + σ _V σ _C + σ _C σ _H )
Thus, the elements derived as a quadratic form by approximating the Hamiltonian to the mean field and the fluctuation field correspond to each energy element, and create discrete energy levels. The values of these energy elements vary depending on the image, but the image group distribution is that the energy element acts as a factor that strongly acts in common among images with a certain sensitivity. It becomes visible by examining it.

In general, as can be seen by taking a color histogram of the image, the momentum distribution has a highly variable characteristic that varies very randomly depending on the image, and even if a statistical average is taken for a group of images with the same sensitivity, a common factor Almost no elements remain. However, when the distribution is observed in a secondary form that is the form of energy, factors that act in common on image groups having the same sensitivity tend to remain. For example, an image that is “killing scenery” has a high probability that the average luminance <V> and the average saturation <C> have a small value at the same time, and the kinetic energy <V><C> shows a small value in common. There are many. However, in some cases, the value of <V> may be larger, in which case <C> is balanced by showing a smaller value. In such a case, if <V> and <C> are individually statistically averaged, those momentum averages tend to disappear as information in the average value of the general image. This is shown below.
Momentum p energy E _n
Hue H Statistical average <V><C>
Luminance V ⇒ <V><V>
Saturation C Residual information <H> σ _V
σ _C σ _C
...
(Random distribution) (Simultaneous alignment)
1) Momentum Examples of the momentum element _pn include the following.
<H>, <V>, <C>, σ _H , σ _V , σ _C
Note that the portion of σ may be treated as an element M _n of angular momentum.

2) The elements E _n of Energy Energy include the following. Energy elements E _n, because different energy elements are calculated, will be represented their state ellipsize the following as a symbol for distinguishing them. The symbol a is used for the α plane, b for the β plane, m for the average value, and s for the standard deviation value.
(α) (α)
amam: <H><H>,<V><V>,<C><C>,
amas: <H> σ _H , <V> σ _V , <C> σ _C ,
asas: σ _H σ _H , σ _V σ _V , σ _C σ _C ,
(α) (β)
ambm: <H><V>,<V><C>,<C><H>,
ambs: <H> σ _V , <V> σ _C , <C> σ _H ,
asbm: σ _H <V>, σ _V <C>, σ _C <H>,
asbs: σ _H σ _V , σ _V σ _C , σ _C σ _H
Regarding the low-order invariants of color, the term with <H> is separated into two elements in consideration of the precautions described below. Therefore, 21 + 5 = 26 types of scalar invariants are derived as energy elements of low-order invariants of color. These are each correspond to discrete energy levels E _n, represents the value itself of E _n. This is a macroscopic physical quantity that has been reduced in the lower-order subsystem of the color.

3) Special matters in the case of low-order invariants of color Hue H is represented by neutral N and other hue ring H (≠ N) during Munsell color space conversion. As described in the first embodiment, the distribution function f (H) uses a histogram divided into hue ring histogram bins and N histogram bins. In addition, when calculating the average value in the hue circle, in the same way as defining the starting point of the hue circle when the Chebyshev expansion is performed in the first embodiment, a cut is made at the point where the distribution function is minimum in the hue circle. In the region with a smaller value than that, the value is extended by adding a value of 2π angle to the end point on the maximum value of the hue ring angle to secure the region, and the average Munsell hue value is calculated on that axis To do. If the value exceeds the original Munsell hue value range, the value corresponding to the 2π angle is subtracted and restored.

In order to express the hue circle, <H> is expressed as a complex number and separated into two components to realize the description. Moreover, the ratio which subtracted the neutral part which remains in the hue ring at the time of the absolute value of a complex number at that time is represented.
<H> = (1-pop (N)) exp (2πi <H (≠ N)> / 100)
σ _H calculates the spread width of only the distribution function in the hue circle. Therefore, it remains one component. This intensity is also evaluated by multiplying the frequency ratio remaining in the hue ring in conjunction with the absolute value of <H>. That is, when everything flows in neutral, σ _H is always defined to be zero.
σ _H = (1-pop (N)) σ _{H (≠ N)}
<Lower-order invariants of texture>
As the edge surface representing the texture, the integrated edge surface described in the second embodiment, in which each Munsell HVC surface is subjected to multi-resolution conversion and only high-frequency subband images are integrated by inverse conversion, is used. The edge strength of the edge surface is schematically expressed as ΔH, ΔV, ΔC using a Laplacian symbol Δ.
Now consider that the values of the edge strengths ΔH, ΔV, and ΔC on the HVC surface represent the momentum.
The model Hamiltonian is constructed as follows.
H = (△ H + △ V + △ C) ^ 2
This represents the kinetic energy representing the intensity of the value of the color edge component, or the energy of the field.

Similar to when the energy eigenvalue E _n lower order invariants color determined by the mean-field approximation. The average value and the standard deviation value of the momentum are obtained from the distribution function f (p) representing the momentum distribution. That is, it is obtained from the histograms f (ΔH), f (ΔV), and f (ΔC) of the edge intensity of ΔH, ΔV, and ΔC.
The low-order invariant of the texture is obtained by the same procedure as that for the low-order invariant of color.
1) Momentum Examples of the momentum element _pn include the following.
<△ H>, <△ V>, <△ C>, σ _{△ H} , σ _{△ V} , σ _{△ C}
Note that the portion of σ may be treated as an element M _n of angular momentum.

2) The elements E _n of Energy Energy include the following. Energy elements E _n, because different energy elements are calculated, will be represented their state ellipsize the following as a symbol for distinguishing them. The symbol a is used for the α plane, b for the β plane, m for the average value, and s for the standard deviation value.
(α) (α)
amam: <△ H><△H>,<△V><△V>,<△C><△C>,
amas: <△ H> σ _{△ H} , <△ V> σ _{△ V} , <△ C> σ _{△ C} ,
asas: σ _{△ H} σ _{△ H} , σ _{△ V} σ _{△ V} , σ _{△ C} σ _{△ C} ,
(α) (β)
ambm: <△ H><△V>,<△V><△C>,<△C><△H>,
ambs: <△ H> σ _{△ V} , <△ V> σ _{△ C} , <△ C> σ _{△ H} ,
asbm: σ _{△ H} <△ V>, σ _{△ V} <△ C>, σ _{△ C} <△ H>,
asbs: σ _{△ H} σ _{△ V} , σ _{△ V} σ _{△ C} , σ _{△ C} σ _{△ H} ,
Twenty-one types of scalar invariants are derived as the energy elements of the low-order invariants of the texture. These are each correspond to discrete energy levels E _n, represents the value itself of E _n. This is a macroscopic physical quantity that has been reduced in the lower order subsystem of the texture.

3) Special matters in the case of low-order invariants of texture The edge component of the hue plane shall take the edge of the color plane displayed with a value where the Munsell value is always 0 as the origin of the hue circle. This is because, rather than experimentally changing the cut point of the hue ring to a point where the distribution function is minimized for each image, it is better to observe the hue plane in a fixed manner at the origin of the same Munsell value as the iridescent spectrum distribution. This is because it has been found preferable as an edge. Further, neutral is not treated separately, but is a hue plane distributed at any point of the hue circle. Therefore, the neutral component behaves like random noise on the hue plane.

分布関数から力学的不変量である運動量、角運動量、エネルギーを導くにあたり、分布関数の形状を記述するのに独立な形状評価をしうる成分を可能な限り導き出す。すなわち、エネルギーや角運動量を構築するに当たり、分布関数f(x)を軸反転したf(-x)も考察対象に加える。色の場合は、軸反転することの物理的意味は階調反転することに相当する。そのために角運動量の概念を導入する。すなわち、基底関数群を偶関数と奇関数の性質の違いによって部分群に分け、各々の部分群に角運動量量子数を割り当てる。 <Higher-order invariants of color>
The Chebyshev expansion of the color distribution function. The variable x takes each value of H, V, and C.

When deriving momentum, angular momentum, and energy, which are mechanical invariants, from the distribution function, components that can be independently evaluated for describing the shape of the distribution function are derived as much as possible. That is, in constructing energy and angular momentum, f (−x) obtained by inverting the axis of the distribution function f (x) is added to the consideration. In the case of color, the physical meaning of axis inversion corresponds to gradation inversion. For this purpose, the concept of angular momentum is introduced. That is, the basis function group is divided into subgroups according to the difference between the even function and the odd function, and an angular momentum quantum number is assigned to each subgroup.

  The Chebyshev basis is an even-numbered basis satisfying the relationship ψ (-x) = ψ (x), and the odd-numbered basis an odd function satisfying the relationship ψ (-x) =-ψ (x) . Therefore, the angular momentum quantum number l = 0 is assigned to the even function group, and the angular momentum quantum number l = 1 is assigned to the odd function group. By performing axis inversion x → −x, the basis functions of odd-numbered angular momentum quantum numbers are inverted in sign, and the basis functions of even-numbered angular momentum quantum numbers are unchanged. Such a property relating to the axis inversion of the wave function of the angular momentum unit is called parity in quantum mechanics. Even-function basis functions have even parity and are invariant to axis inversion, and odd-function basis functions have odd parity and invert the sign with respect to axis inversion.

  If this angular momentum quantum number is considered to describe a spin system, the color describes a system having a spin angular momentum quantum number s = 1. The Chebyshev function is defined only by a single series expansion, the even function and the odd function describe the states of spin 0 and spin 1, respectively, and a state in which the parity of the state of spin 1 is inverted by axis inversion can be generated. The Chebyshev function is suitable for describing a spin-1 system that is a Bose particle.

In considering the independent components, first, the momentum element p _n , the angular momentum element M _n , and the energy element E _n when the axis of the angular momentum describing the system is not reversed are constructed, and then each of them is constructed. The element of about twice the angular momentum inversion is derived. At that time, when all elements are linearly combined, the linear combination coefficient is derived from the viewpoint of describing a meaningful independent component. That is, when only the sign of an element changes, it does not make sense because the same system is described if the sign of the linear combination coefficient is changed. In general, momentum falls into this category, but angular momentum and energy can be derived as separate components.
To give a specific example, the expansion coefficient number 2N is set to 200. That is, N = 100.

1) Momentum Examples of the momentum element _pn include the following.
c _n ^(α) where (α) = H, V, C.
When 2N = 200, there are three momentum elements, so 200x3 = 600.

2) Angular momentum In the following, when c _n is divided into subgroups of angular momentum units, the expansion coefficient of angular momentum quantum number l = 0 is c _0n , and the expansion coefficient of angular momentum quantum number l = 1 is c _1n. It will be expressed as Therefore, the number of elements of the expansion coefficient is divided into N half. The element numbers of the subgroups shall be counted by n = 1, 2, ..., N.
Examples of the angular momentum element M _n include the following.
0 * (c ₀₁ ^(α) + c ₀₂ ^(α) + ... + c _0N ^(α) ) + 1 * (c ₁₁ ^(α) + c ₁₂ ^(α) + ... + c _1N ^(α) )
= (c ₁₁ ^(α) + c ₁₂ ^(α) + ... + c _1N ^(α) ) where (α) = H, V, C.

When the angular momentum exists only up to l = 1, there is only one independent component as described above. Because the axis of angular momentum is reversed
0 * (c ₀₁ ^(α) + c ₀₂ ^(α) + ... + c _0N ^(α) ) -1 * (c ₁₁ ^(α) + c ₁₂ ^(α) + ... + c _1N ^(α) ) =-(c ₁₁ ^(α) + c ₁₂ ^(α) + ... + c _1N ^(α) ) only describes the same system as above.

As described above, the linear sum of the expansion coefficient of the odd function can be a macroscopic index for evaluating the asymmetry of the distribution function. The classical definition of angular momentum is M = rxp, but in contrast, the above definition takes the product in a coordinate space where the angular momentum quantum number of the Hilbert space coordinates is the distance from the origin for the partial sum of the momentum. Thus, the moment of the distribution function is described.
Since there are three elements of angular momentum, 1x3 = 3.

3) Energy Build the kinetic energy by taking the product of the momentum of the α and β planes. The momentum product c _m * c _n creates a matrix called the direct product or the Kronecker product in group theory, but the product matrix of two systems represented by the same basis function is a reducible representation, and is expressed in two representations of smaller dimensions. Can be disassembled. That is, it can be decomposed into a matrix representation of a symmetric product and an antisymmetric product,
From the product of the (α) plane and the (α) plane, a symmetrical product energy matrix (i, k) can be constructed.
c _i ^(α) c _k ^(α) + c _k ^(α) c _i ^(α)
Also, an energy matrix (i, k) of a symmetric product and an antisymmetric product can be constructed from the product of the (α) plane and the (β) plane.
c _i ^(α) c _k ^(β) + c _k ^(α) c _i ^(β)
c _i ^(α) c _k ^(β) -c _k ^(α) c _i ^(β)
Each of these is a square matrix arranged vertically and horizontally by the number of basis functions.

In order to calculate an eigenvalue of energy, a diagonal sum, that is, a spur is taken. The steady state energy eigenvalues are purely diagonal sums. This is the energy elements E _n in the case of n = 0 i.e. i = k. In order to calculate the energy eigenvalue of the unsteady state, an extended spur that takes the sum of matrix elements separated from the diagonal component by n = ik is defined. Since it is different from normal spurs, the symbol Sp 'is used. At this time, the number of elements to be spured must be defined to be the same as the number of elements in the diagonal sum, and the basis functions used in one of the products must lick all the basis functions that make up the complete system once. Build in shape. This is a requirement imposed as a rule when decomposing into symmetric and antisymmetric products. For this reason, when the sum operation is performed with matrix elements satisfying n = ik, the elements that protrude at the end of the matrix are assumed to be within the matrix if the size of the matrix or submatrix to be considered is N. In order to satisfy n + N = ik, the diagonal component is defined so that the diagonal sum is taken for the remaining number of elements in the opposite region. Thus the energy elements E _n are sequentially calculated. Specific examples have been described in the first embodiment and the third embodiment.

  There are as many energy elements as one half of the number of expansion coefficients of the basis function for one energy matrix. The reason for the halving is that the component protruding from the matrix is shifted by the number equal to the matrix width and is woven again, so if they are licked in order, the definition is double, so that the effective is reduced to half. That is, by taking an expanded spur, the number of energy elements is reduced to a number equal to half of the number of one-dimensional elements forming each row and column.

  Next, consider the state in which the axis of angular momentum is reversed. The definition of the three energy matrices described above describes the appearance of the table by the product of the surface comprising the (α) plane and the (α) plane, and the (α) plane and the (β) plane. On the other hand, the (−α) plane or the (−β) plane on the back surface can be created by reversing the axis of one distribution function. When making a matrix product, it is possible to depict the back by taking one on the front and the other on the back. This makes it possible to derive independent energy components. That is, by introducing the concept of parity, an independent synthetic basis function is constructed. Parity is a concept that does not appear in classical mechanics.

Here, an energy matrix (i, k) of a symmetric product and an antisymmetric product can be constructed from the product of the (α) plane and the (−α) plane.
c _i ^(α) c _k ^(-α) + c _k ^(α) c _i ^(-α)
c _i ^(α) c _k ^(-α) -c _k ^(α) c _i ^(-α)
Also, an energy matrix (i, k) of a symmetric product and an antisymmetric product can be constructed from the product of the (α) plane and the (−β) plane.
c _i ^(α) c _k ^(-β) + c _k ^(α) c _i ^(-β)
c _i ^(α) c _k ^(-β) -c _k ^(α) c _i ^(-β)
Only if k of each of the elements of the energy matrix c _k ^(-.alpha.) is equivalent to odd ^{_{function, c k (-α) = -c}} k (α) the sign is inverted, When it corresponds to an even function, an energy matrix that does not change the sign is constructed such as c _k ^(−α) = c _k ^(α) . In this way, if an extended spur similar to that of the front matrix is taken in the matrix of the reverse figure in which the sign of some elements is reversed, an independent component can be derived as an energy element that appears individually.

第２番目は主量子数の低い順番から先に並べる方法である。

これらを順番にψ_iと一次元の指標を与え、２次元の行列積として対称積と反対称積ψ_iψ_k±ψ_kψ_iを作ると、第１番目の項と第２番目の項のiとkは対角成分を挟んで交換することになる。これらのエネルギー行列をもう一度、角運動量単位で部分行列に分けて並べると、部分行列内でn=i-kを満たす拡張シュプールをとることと等価になる。ただし、i,kを角運動量量子数と主量子数の２つのインデックスを用いて表した場合、第１番目の並べ方の場合は単純に同じ角運動量量子数の組み合せの部分行列内で主量子数のiとkを交換した形で表せるが、第２番目の並べ方を採用した場合、少しそれとは異なる表現となるので注意を要する。すなわち、同じ距離の非対角位置に存在する部分行列の要素との間でi,kを交換するとき、第１項目のψ_iψ_kのψ_i側がψ_k側よりも大きな角運動量量子数の場合は、n=i-k個だけ右側に来るψの主量子数はψ_i側の主量子数よりも１つだけ繰り上がったk+1として表記しなければならない。その後に、どちらの並べ替えをした場合であっても、要素を交換した２つの部分行列の間で、部分行列単位で要素をまとめる操作を行って部分行列表現する。 There are two ways to arrange the basis functions when subgroups based on angular momentum can be defined when taking the extended spur of the energy matrix shown in the table or the extended spur of the energy matrix of the back. That is, the first is a method in which the angular momentum quantum numbers are arranged in ascending order.

The second is a method in which the main quantum numbers are arranged in ascending order.

Given these in turn, ψ _i and a one-dimensional index, and creating a symmetric product and an anti-symmetric product ψ _i ψ _k ± ψ _k ψ _i as a two-dimensional matrix product, the first and second terms I and k are exchanged across a diagonal component. If these energy matrices are divided into sub-matrices in units of angular momentum, they are equivalent to taking an extended spur that satisfies n = ik in the sub-matrix. However, when i and k are expressed using two indexes of angular momentum quantum number and main quantum number, in the case of the first arrangement, the main quantum number is simply within the submatrix of the same combination of angular momentum quantum numbers. However, if the second arrangement is adopted, the expression will be slightly different from this, so care must be taken. That is, when exchanging i and k with submatrix elements existing at off-diagonal positions at the same distance, the angular momentum quantum number in which the ψ _i side of ψ _i ψ _k of the first item is larger than the ψ _k side In this case, the main quantum number of ψ coming to the right by n = ik pieces must be expressed as k + 1 which is one higher than the main quantum number on the ψ _i side. Thereafter, regardless of which rearrangement is performed, an operation for grouping elements in units of partial matrix is performed between the two partial matrices whose elements have been exchanged, and the partial matrix is expressed.

  Experimentally, the first arrangement is better when creating the product of the (α) and (α) planes, and the second arrangement when creating the product of the (α) and (β) planes. Turned out to be excellent. This is based on the premise that the first arrangement can be treated as an independent system between different angular momentum quantum numbers, and that it is satisfied by the expansion of the basis function of the completely orthogonal system in the same plane. means. On the other hand, the second way of arrangement shows that even if it is between different angular momentum quantum numbers, it is possible to describe how the basis functions between the adjacent main quantum numbers are more closely related. This is probably because the HVC surface does not describe a completely independent system.

As a method for combining such sub-matrices, it is necessary to take a sum between the sub-matrices so that the extended spurs lick all the basis functions forming the complete system. That is, when the momentum element is divided into two subgroups consisting of systems of angular momentum quantum numbers 0 and 1, the energy matrix is a mixture of two submatrices representing the eigenstates of angular momentum and angular momentum. Can be divided into two sub-matrices representing. At this time, there are two ways of taking the extended spurs that lick to form a complete system: connecting two spurs of diagonal sub-matrices and connecting two spurs of non-diagonal sub-matrices. Thus, the physical meaning becomes clearer when the energy matrix is divided into units of angular momentum.
This state is schematically shown as follows.

When the same suffix appears, it shall be summed up.
A specific example of a submatrix sum for actually dividing an energy matrix into submatrices of angular momentum to form an extended spur will be shown. The following is expressed in the form of a composite system of two basis functions forming a submatrix, and the sum of the submatrix and submatrix is also shown. Each submatrix can be obtained by simply replacing these basis functions ψ _ik with c _ik . Therefore, when each submatrix is defined by (i, k), in _order to calculate the energy element En, a spur is performed on the matrix elements satisfying n = ik in each submatrix.

In subgroup units to perform spur the energy elements E _n, because different energy elements are calculated, it will be represented their state ellipsize the following as a symbol for distinguishing them. a for the α plane, b for the β plane, 00 symbol for the combination of (l, l ') = (0,0) + (1,1) where the angular momentum is an eigenstate, and the angular momentum is mixed Use the 01 symbol for the (l, l ') = (0,1) + (1,0) combination of states, the positive p symbol for symmetric matrices, and the negative m symbol for antisymmetric matrices. The symbol e is used when the angular momentum coordinate axis is in the standard state, and the symbol i is used when the angular momentum coordinate axis is inverted.
Regarding the angular momentum coordinate axis, the standard state and the inverted state will be described simultaneously using ± symbols. Since the axis inversion operation of the angular momentum for describing the back side is performed only on one basis function constituting the matrix, only the odd function on one color plane side is inverted in sign. Components that disappear through these operations are not described.

(α) (α)
a0a0p, e / i: (ψ _0i ^(α) ψ _0k ^(α) + ψ _0k ^(α) ψ _0i ^(α) ) ± (ψ _1i ^(α) ψ _1k ^(α) + ψ _1k ^(α) ψ _1i ^{( α)} )
a0a1p, e: (ψ _0i ^(α) ψ _1k ^(α) + ψ _0k ^(α) ψ _1i ^(α) ) + (ψ _1i ^(α) ψ _0k ^(α) + ψ _1k ^(α) ψ _0i ^(α) )
a0a1m, i:-(ψ _0i ^(α) ψ _1k ^(α) -ψ _0k ^(α) ψ _1i ^(α) ) + (ψ _1i ^(α) ψ _0k ^(α) -ψ _1k ^(α) ψ _0i ^{(α )} ) (i ≠ k)
(α) (β)
a0b0p, e / i: (ψ _0i ^(α) ψ _0k ^(β) + ψ _0k ^(α) ψ _0i ^(β) ) ± (ψ _1i ^(α) ψ _1k ^(β) + ψ _1k ^(α) ψ _1i ^{( β)} )
a0b0m, e / i: (ψ _0i ^(α) ψ _0k ^(β) -ψ _0k ^(α) ψ _0i ^(β) ) ± (ψ _1i ^(α) ψ _1k ^(β) -ψ _1k ^(α) ψ _1i ^{( β)} ) (i ≠ k)
a0b1p, e / i: ± (ψ _0i ^(α) ψ _1k ^(β) + ψ _{0, k + 1} ^(α) ψ _1i ^(β) ) + (ψ _1i ^(α) ψ _{0, k + 1} ^(β) + ψ _1k ^(α) ψ _0i ^(β) )
a0b1m, e / i: ± (ψ _0i ^(α) ψ _1k ^(β) -ψ _{0, k + 1} ^(α) ψ _1i ^(β) ) + (ψ _1i ^(α) ψ _{0, k + 1} ^(β) -ψ _1k ^(α) ψ _0i ^(β) )
The number of energy elements E _n coming defined from consisting of partial matrix 100x100 closed with respective brackets, the spurs from each type when the 2N = 200 is 50. The number of types defined above has four types of products between the same color surfaces and eight types of products between the different color surfaces. Further, there are three types of (α) (α) taking HH, VV, and CC, and there are three types of (α) (β) taking HV, VC, and CH. Therefore, the number of energy elements of the system described by the angular momentum of 0 and 1 is (3 × 4 + 3 × 8) × 50 = 36 × 50 = 1800. As for the energy bands, for each of the 36 types, a diagram having 50 energy levels can be drawn.

  As a result of actually experimenting with images, the energy matrix is displayed as a partial matrix in the unit of angular momentum as described above, and the submatrix of the eigenstate with the same angular momentum quantum number is different in the diagonal position. If the submatrix of the mixed state (hybridization term) with the angular motion quantum number appears in the off-diagonal position, the results consistent with the physical meaning can be obtained. In other words, when comparing the image alignment for energy elements made from diagonal submatrices and the image alignment for energy elements made from off-diagonal submatrices, clouds emerge from the mountains in the mixed state of off-diagonal components. It was found that it has the ability to capture very dynamic pictures such as scenes.

<Higher-order invariants of texture>
Expand the texture distribution function into a spherical Bessel. The spherical Bessel function is extended to the negative region and expanded to a double series of roots and orders that form a complete system. Since the spherical Bessel function represents a radial wave function, the expansion relating to the root corresponds to the principal quantum number n, and the expansion relating to the order corresponds to the orientation quantum number (orbital angular momentum quantum number) l. Consider two cases, when the expansion coefficient of the orientation quantum number is l = 0,1, and when l = 0,1,2,3. In atomic physics, the orbits corresponding to l = 0,1,2,3 are given different names and are called s-orbital, p-orbital, d-orbital, and f-orbital in that order. These orbits form the periodic table of elements, the d orbital corresponds to the transition metal, and the f orbit corresponds to the description of the electron system of the lanthanoid and actinoid elements. Considering the correspondence with the electronic system describing physical properties, it is considered sufficient to expand to the f orbit even in the image system describing sensibility. The variable x takes the values of ΔH, ΔV, and ΔC.

In case of expansion by s, p orbitals

Case of expansion by s, p, d, f orbitals

Here, (α) is used to represent the difference in color plane, but α _ln in the spherical Bessel function is used to represent the root position. In general, the position of the zero point cannot be expressed analytically, but Smirnov's "Higher Mathematics Curriculum" contains an approximate expression for the zero of the Bessel function. Then, by deducting it into an approximate expression to represent the zero point of the spherical Bessel function and introducing a correction factor of 1/2 with respect to the second term without changing the first term, an error within 3% at all zeros Since it has been found that the expansion coefficient can be approximated by the following equation, the following equation is used to calculate the expansion coefficient over the p orbit.

  When deriving momentum, angular momentum, and energy, which are mechanical invariants, from the distribution function, components that can be independently evaluated to describe the shape of the distribution function are derived as much as possible. That is, as in the case of color, in constructing energy and angular momentum, f (−x) obtained by axially inverting the distribution function f (x) is also added to the consideration. In the case of texture, the physical meaning of axis inversion corresponds to edge code inversion. The axis inversion operation corresponds to an operation to invert the parity of angular momentum. In the spherical Bessel function, the angular momentum quantum numbers are classified into subgroups based on the difference between the properties of the even function and the odd function, and an orientation quantum number is assigned to each subgroup.

  A spherical Bessel basis has an even-numbered orientation quantum number that is an even function and satisfies the relationship ψ (-x) = ψ (x), and an odd-numbered orientation quantum number basis that is an odd function ψ (-x ) =-ψ (x). By performing the axis inversion x → −x, the basis function of the odd-numbered azimuth quantum number is inverted, and the basis function of the even-numbered azimuth quantum number has the same sign.

  Therefore, the wave function of the s and d orbits has an even parity, and the wave function of the p and f orbits has an odd parity. In general, a basis function group represented by a spherical Bessel function has a parity represented by (−1) ^ l using an angular momentum quantum number l with respect to inversion of coordinate axes.

As in the previous case, when considering the independent components, first, the momentum element p _n , the angular momentum element M _n , and the energy element E _n when the axis of the angular momentum describing the system is not reversed are first constructed. Thus, each of them leads to an element of about twice the axis of angular momentum inverted. In order to give a specific example, the number N of expansion coefficients is set to 100.

1) Momentum Examples of the momentum element _pn include the following.
c _ln ^(α) where (α) = H, V, C.
In the case of expansion by the s, p orbit, there are 3 momentum elements when N = 100, so 2x100x3 = 600. In the case of expansion by the s, p, d, f orbit, when N = 100, there are three elements of momentum, so 4 × 100 × 3 = 1200.

2) Angular momentum Examples of the angular momentum element M _n include the following.
(In case of expansion by s, p orbit)
This is exactly the same as in the case of color Chebyshev expansion. That is,
(c ₁₁ ^(α) + c ₁₂ ^(α) + ... + c _1N ^(α) ) where (α) = H, V, C.
Since there are three elements of angular momentum, 1x3 = 3.
(In case of expansion by s, p, d, f orbit)
1 (c ₁₁ ^(α) + c ₁₂ ^(α) + ... + c _1N ^(α) ) +2 (c ₂₁ ^(α) + c ₂₂ ^(α) + ... + c _2N ^(α) ) + 3 (c ₃₁ ^(α) + c ₃₂ ^(α) + ... + c _3N ^(α) )
When the axis of angular momentum is reversed, another independent component appears.
-1 (c ₁₁ ^(α) + c ₁₂ ^(α) + ... + c _1N ^(α) ) +2 (c ₂₁ ^(α) + c ₂₂ ^(α) + ... + c _2N ^(α) ) -3 (c ₃₁ ^(α) + c ₃₂ ^(α) + ... + c _3N ^(α) )
Since there are three elements of angular momentum, 2x3 = 6.

  In this way, the linear sum of the expansion coefficient of the odd function gives a macroscopic index for evaluating how far the asymmetry of the distribution function extends to show its properties, and further, the expansion coefficient of the even function The linear sum gives a macroscopic index for evaluating the extent to which the distribution function spreads in a tail-extracting manner, and preserves the overall spread property as angular momentum It can be a macroscopic physical quantity.

3) As the elements E _n of Energy Energy include the following.
(In case of expansion by s, p orbit)
This is exactly the same as in the case of color Chebyshev expansion. Therefore, submatrix closed with respective brackets when N = 100 from the consisting of 100x100, the number of energy elements E _n coming defined by spurs from each type is 50. Therefore, just like the case of color, the number of energy elements of the system described by the angular momentum of 0 and 1 is (3 × 4 + 3 × 8) × 50 = 36 × 50 = 1800. As for the energy bands, for each of the 36 types, a diagram having 50 energy levels can be drawn.

(In case of expansion by s, p, d, f orbit)
There are the following four cases in which the energy submatrix of angular momentum units is connected diagonally to form a complete system. The first bundling method represents the intrinsic state of angular momentum. The second to fourth bundling methods represent a mixed state of angular momentum. The mixed state is a state in which the angular momentum has a hybrid orbit, for example, a subgroup called sd hybridization is created between sd. 17A to 17D are matrix diagrams showing how the following four extended spurs are configured. These are defined to include the case of only s, p orbitals. That is, when expansion by d, f orbit is omitted, the same expression as expansion by s, p orbit is obtained.
s ² + p ² + d ² + f ²
sp + ps + df + fd
sd + ds + pf + fp
sf + fs + pd + dp
A specific example of a submatrix sum for actually dividing an energy matrix into submatrices of angular momentum to form an extended spur will be shown. For the combination of (l, l ') = (0,0) + (1,1) + (2,2) + (3,3) where the angular momentum is an eigenstate, the 00 symbol is mixed and the angular momentum is mixed For the combination of (l, l ') = (0,1) + (1,0) + (2,3) + (3,2) in the state, use the 01 symbol and (l , l ') = (0,2) + (2,0) + (1,3) + (3,1), use the 02 symbol, and (l, l') with a mixed angular momentum The 03 symbol is used for the combination of = (0,3) + (3,0) + (1,2) + (2,1).

(α) (α)
a0a0p, e / i: (ψ _0i ^(α) ψ _0k ^(α) + ψ _0k ^(α) ψ _0i ^(α) ) ± (ψ _1i ^(α) ψ _1k ^(α) + ψ _1k ^(α) ψ _1i ^{( α)} ) + (ψ _2i ^(α) ψ _2k ^(α) + ψ _2k ^(α) ψ _2i ^(α) ) ± (ψ _3i ^(α) ψ _3k ^(α) + ψ _3k ^(α) ψ _3i ^(α) )
a0a1p, e: (ψ _0i ^(α) ψ _1k ^(α) + ψ _0k ^(α) ψ _1i ^(α) ) + (ψ _1i ^(α) ψ _0k ^(α) + ψ _1k ^(α) ψ _0i ^(α) ) + (ψ _2i ^(α) ψ _3k ^(α) + ψ _2k ^(α) ψ _3i ^(α) ) + (ψ _3i ^(α) ψ _2k ^(α) + ψ _3k ^(α) ψ _2i ^(α) )
a0a1m, i:-(ψ _0i ^(α) ψ _1k ^(α) -ψ _0k ^(α) ψ _1i ^(α) ) + (ψ _1i ^(α) ψ _0k ^(α) -ψ _1k ^(α) ψ _0i ^{(α )} )-(ψ _2i ^(α) ψ _3k ^(α) -ψ _2k ^(α) ψ _3i ^(α) ) + (ψ _3i ^(α) ψ _2k ^(α) -ψ _3k ^(α) ψ _2i ^(α) ) (i ≠ k)
a0a2p, e / i: (ψ _0i ^(α) ψ _2k ^(α) + ψ _0k ^(α) ψ _2i ^(α) ) + (ψ _2i ^(α) ψ _0k ^(α) + ψ _2k ^(α) ψ _0i ^{( α)} ) ± (ψ _1i ^(α) ψ _3k ^(α) + ψ _1k ^(α) ψ _3i ^(α) ) ± (ψ _3i ^(α) ψ _1k ^(α) + ψ _3k ^(α) ψ _1i ^(α) )
a0a3p, e: (ψ _0i ^(α) ψ _3k ^(α) + ψ _0k ^(α) ψ _3i ^(α) ) + (ψ _3i ^(α) ψ _0k ^(α) + ψ _3k ^(α) ψ _0i ^(α) ) + (ψ _1i ^(α) ψ _2k ^(α) + ψ _1k ^(α) ψ _2i ^(α) ) + (ψ _2i ^(α) ψ _1k ^(α) + ψ _2k ^(α) ψ _1i ^(α) )
a0a3m, i:-(ψ _0i ^(α) ψ _3k ^(α) -ψ _0k ^(α) ψ _3i ^(α) ) + (ψ _3i ^(α) ψ _0k ^(α) -ψ _3k ^(α) ψ _0i ^{(α )} ) + (ψ _1i ^(α) ψ _2k ^(α) -ψ _1k ^(α) ψ _2i ^(α) )-(ψ _2i ^(α) ψ _1k ^(α) -ψ _2k ^(α) ψ _1i ^(α) ) (i ≠ k)
(α) (β)
a0b0p, e / i: (ψ _0i ^(α) ψ _0k ^(β) + ψ _0k ^(α) ψ _0i ^(β) ) ± (ψ _1i ^(α) ψ _1k ^(β) + ψ _1k ^(α) ψ _1i ^{( β)} ) + (ψ _2i ^(α) ψ _2k ^(β) + ψ _2k ^(α) ψ _2i ^(β) ) ± (ψ _3i ^(α) ψ _3k ^(β) + ψ _3k ^(α) ψ _3i ^(β) )
a0b0m, e / i: (ψ _0i ^(α) ψ _0k ^(β) -ψ _0k ^(α) ψ _0i ^(β) ) ± (ψ _1i ^(α) ψ _1k ^(β) -ψ _1k ^(α) ψ _1i ^{( β)} ) + (ψ _2i ^(α) ψ _2k ^(β) -ψ _2k ^(α) ψ _2i ^(β) ) ± (ψ _3i ^(α) ψ _3k ^(β) -ψ _3k ^(α) ψ _3i ^(β) ) (i ≠ k)
a0b1p, e / i: ± (ψ _0i ^(α) ψ _1k ^(β) + ψ _{0, k + 1} ^(α) ψ _1i ^(β) ) + (ψ _1i ^(α) ψ _{0, k + 1} ^(β) + ψ _1k ^(α) ψ _0i ^(β) ) ± (ψ _2i ^(α) ψ _3k ^(β) + ψ _{2, k + 1} ^(α) ψ _3i ^(β) ) + (ψ _3i ^(α) ψ _{2, k + 1} ^(β) + ψ _3k ^(α) ψ _2i ^(β) )
a0b1m, e / i: ± (ψ _0i ^(α) ψ _1k ^(β) -ψ _{0, k + 1} ^(α) ψ _1i ^(β) ) + (ψ _1i ^(α) ψ _{0, k + 1} ^(β) -ψ _1k ^(α) ψ _0i ^(β) ) ± (ψ _2i ^(α) ψ _3k ^(β) -ψ _{2, k + 1} ^(α) ψ _3i ^(β) ) + (ψ _3i ^(α) ψ _{2, k + 1} ^(β) -ψ _3k ^(α) ψ _2i ^(β) )
a0b2p, e / i: (ψ _0i ^(α) ψ _2k ^(β) + ψ _{0, k + 1} ^(α) ψ _2i ^(β) ) + (ψ _2i ^(α) ψ _{0, k + 1} ^(β) + ψ _2k ^(α) ψ _0i ^(β) ) ± (ψ _1i ^(α) ψ _3k ^(β) + ψ _{1, k + 1} ^(α) ψ _3i ^(β) ) ± (ψ _3i ^(α) ψ _{1, k +1} ^(β) + ψ _3k ^(α) ψ _1i ^(β) )
a0b2m, e / i: (ψ _0i ^(α) ψ _2k ^(β) -ψ _{0, k + 1} ^(α) ψ _2i ^(β) ) + (ψ _2i ^(α) ψ _{0, k + 1} ^(β) - ψ _2k ^(α) ψ _0i ^(β) ) ± (ψ _1i ^(α) ψ _3k ^(β) -ψ _{1, k + 1} ^(α) ψ _3i ^(β) ) ± (ψ _3i ^(α) ψ _{1, k +1} ^(β) -ψ _3k ^(α) ψ _1i ^(β) )
a0b3p, e / i: ± (ψ _0i ^(α) ψ _3k ^(β) + ψ _{0, k + 1} ^(α) ψ _3i ^(β) ) + (ψ _3i ^(α) ψ _{0, k + 1} ^(β) + ψ _3k ^(α) ψ _0i ^(β) ) + (ψ _1i ^(α) ψ _2k ^(β) + ψ _{1, k + 1} ^(α) ψ _2i ^(β) ) ± (ψ _2i ^(α) ψ _{1, k + 1} ^(β) + ψ _2k ^(α) ψ _1i ^(β) )
a0b3m, e / i: ± (ψ _0i ^(α) ψ _3k ^(β) -ψ _{0, k + 1} ^(α) ψ _3i ^(β) ) + (ψ _3i ^(α) ψ _{0, k + 1} ^(β) -ψ _3k ^(α) ψ _0i ^(β) ) + (ψ _1i ^(α) ψ _2k ^(β) -ψ _{1, k + 1} ^(α) ψ _2i ^(β) ) ± (ψ _2i ^(α) ψ _{1, k + 1} ^(β) -ψ _2k ^(α) ψ _1i ^(β) )
From consisting of 100x100 is closed submatrix in each parenthesis when N = 100, the number of energy elements E _n coming defined by spurs from each type is 50. The number of types defined above includes 8 types of products between the same color surfaces and 16 types of products between different color surfaces. Further, there are three types of (α) (α) taking HH, VV, and CC, and there are three types of (α) (β) taking HV, VC, and CH. Therefore, the number of energy elements of the system described by the angular momentum of 0, 1, 2, 3 is (3x8 + 3x16) x50 = 72x50 = 3600. As energy bands, 72 types can each be drawn with 50 energy levels.

<Correction of Hamiltonian relativistic effect>
1) Combination energy of higher order system of color and texture It was assumed that the color is a spin coordinate system and the texture is a radial position coordinate system. In the non-relativistic description, the spin coordinate system and the position coordinate system are independent, and the spin angular momentum and the orbital angular momentum individually act as conserved quantities, but in the relativistic description, the spin coordinate and the position coordinate cannot be distinguished. The influence of relativistic effects can be incorporated to some extent by adding spin-orbit interaction energy to the non-relativistic Hamiltonian, even without spinor description (see reference F2). Therefore, it is possible to construct the spin angular momentum S defined by the color of the higher-order system, the energy element E _n by the inner product of the orbital angular momentum L defined order based texture. However, for the angular momentum M defined so far in each subsystem, the symbol S is used for the higher-order color subsystem and L is used for the higher-order texture subsystem.

The above definition when (α) = H, V, and C are assumed to be independent is the inner product of the (α) plane and the (α) plane, and the energy component when the axial inversion of angular momentum is not considered There is only one number. However, since HVC has non-independent aspects, it is generally the inner product of the (α) plane and the (β) plane, and the following three energy elements can be defined.

Even when the axis of angular momentum is reversed, another L ′ vector and S ′ vector (= S vector) may be defined and the next Hamiltonian added. The number of energy elements is doubled to 6 as described above.

2) Combination energy of low-order system of color and texture The spread σ _H , σ _V , and σ _C of the momentum introduced in the low-order system of color in the position coordinate direction can also be regarded as angular momentum. Similarly spread sigma _{△ H} of the introduced momentum lower order system texture, σ _{△ V,} σ _{△ C} may also be regarded as angular momentum. Therefore, the energy of spin-orbit interaction can be defined as in the case of higher order systems.

  In the low-order system, even if the axis of angular momentum is inverted, it only represents the same system as the original, so the number of energy elements is three.

[Literature F2] Landau and Lifshitz, Course of Theoretical Physics, Volume 3 "Quantum Mechanics (Non-Relativistic Theory)," (Third revised edition, 1977), Chapter 10 "The Atom," Section 72 "Fine Structure of Atomic levels. "
<Addition of internal energy of rotation>
1) Rotational energy of higher-order system of color In the same manner as the above-described spin-orbit interaction, a spin-spin interaction corresponding to the rotational energy of the spin system can be defined.

The following two energy elements can be defined as a combination when (α) = H, V, and C are assumed not to be independent.

2) Rotational energy of low-order color system The low-order system has already been introduced as a product of σ and σ, which is the quadratic form of the fluctuation term of the model Hamiltonian.
3) Rotational energy of higher order system of texture In the same way as the above-mentioned spin-spin interaction, rotational energy by the orbital angular momentum of the coordinate system can be defined.

The following two energy elements can be defined as a combination when (α) = H, V, and C are assumed not to be independent.

When s, p, d, f are expanded, an angular momentum L ′ vector independent of the coordinate axis inversion is defined. As a method of this combination, in addition to the case where one angular momentum is reversed, the case where both angular momentums are reversed is an independent energy element. Therefore, the number of energy elements is three times six.

4) Rotational energy of the low-order system of the texture As with the low-order system of the color, it has already been introduced.

<Linear model of sensitivity>
Since all the mechanical invariants defined in this way have additive properties, they can be described by connecting them all with linear combinations without dividing the world between different dimensional spaces called projections onto subsystems. That is, by constructing a mechanical invariant, it is possible to place a sub-system between the main axes of sensibility and the lower-order and higher-order projections in the common earth.
Differences in physical constants such as Planck's constant h defined differently for each subsystem are included as linear coupling coefficients that scale the entire subsystem. That is, the sensitivity can be described as a linear model.

  The intensity Qi of the impression level of a sensitivity related to a certain adjective i is expressed by a linear sum of individual energy elements emitted by each image. The size and sign of the energy element that is regarded as important for each adjective is different, and it appears in the form of a linear coupling coefficient as a property that characterizes the adjective. In other words, mechanical invariants, which are image features, are physical quantities converted from purely microscopic image information to macroscopic image information, and what needs to be modeled is aggregated in the linear combination coefficient. The The prior learning of the model for determining the linear combination coefficient is also based on the statistical average of the statistical group of the image group, following the basic concept of statistical physics.

Qi = α ₁ F ₁ + α ₂ F ₂ + ... (color)
+ β ₁ G ₁ + β ₂ G ₂ + ... (texture)
+ γ ₁ H ₁ + γ ₂ H ₂ + ... (Composition)
In the pyramid type hierarchical structure, the factor structure once modeled in each layer can be added with other factors as it is. The only point that needs to be changed is the scaling connecting the subsystems, and the ratios within the subsystems, that is, the factor structure is unchanged.

  Furthermore, the important point of the linear model is that each subsystem is described with additive features, so the number of combined features when integrating subsystems such as inter-spindle synthesis is the most reduced. It is expressed by a simple simple sum. The number is on the order of 10,000, which is about the same as the total number of sensitivity adjectives.

  Not only the linear combination of energy but also the remaining additive mechanical invariants, momentum and angular momentum, can be linearly combined with Qi. At that time, the linear coupling coefficient also plays a role in matching energy and physical dimensions. Therefore, when all the mechanical invariants are used, after determining the factor structure of the linear coupling coefficient in energy unit, momentum unit, and angular momentum unit during model learning, the coupling coefficient for scaling the whole between the three It is necessary to decide again. In addition to matching the dimensions, this also determines which invariants play an important role among the three. In general, the role of momentum and angular momentum decreases when statistical averaging is performed on many model images.

  In the next section, we will limit it to energy only to make the story easier to understand.

<State density and adjective model>
If macroscopic energy is derived from image information and has a value at a certain energy level, the energy at that energy level is emitted by that amount, and a state exists at that energy level. Means. If the value of an energy element at an energy level is zero, the image does not emit that energy element at all, and no state exists at that level. The energy value may take a negative value.

  For an adjective i, if the value of the product of the energy element and the linear coupling coefficient acts positively, it acts positively on the adjective i for that element, and conversely if the product value is negative Has the effect of. When the sum of the inner products of a distribution diagram of energy elements derived from a single image, that is, a diagram representing the existence of energy level states and a model coefficient distribution diagram of linear combination, has a positive value, It means that the image has the impression of the adjective i by that value. On the other hand, when the image has a negative value, it can be said that the image emits an impression in the direction opposite to that of the adjective i. Therefore, the sum Qi of the inner products of the linear combination coefficient of the model adjective and the energy element represents the adjective energy.

となるように行う。しかし、状態密度を図示するときや、実際に状態密度と線形結合係数との間で線形結合をとるときの状態密度の表記法としては、このように２つの部分に分けなくてもよい。すなわち、エネルギー値の符号をそのまま採用して、その大きさとしてρ₊(E)ないしはρ_-(E)の該当する方の値を使ってρ(E)と表しておけば、記号を分けなくてもよいのである。また、線形結合係数の方も規格化しておくと下式のようになる。

When the energy element distribution map is normalized by the number of states of the entire energy system, a state density function representing the existence probability of the state is represented. Those having a positive energy value have a positive existence probability in the positive state ρ ₊ (E). On the other hand, the negative state ρ ₋ (E) also has a positive existence probability for those having a negative energy value. Standardization is

To do so. However, the notation of the state density when the state density is illustrated or when the linear combination between the state density and the linear coupling coefficient is actually taken may not be divided into two parts in this way. That is, adopts the sign of the energy value as it, [rho ₊ as its size (E) or [rho _- if expressed by using the appropriate value towards that of (E) [rho and (E), not divided into symbols It may be. If the linear combination coefficient is also normalized, the following equation is obtained.

Here, <> in the denominator represents a statistical average of images that are models of general images, and <> _i represents a statistical average of all adjectives prepared by the image search system. When the same suffix appears for n here, the sum is taken for them.

  In this way, if the norm of the denominator is standardized using the statistical average of the image group and adjective group over the possible range, the adjective impression of how strong an image is with respect to the average image It is possible to quantify whether or not it is emitting on an absolute basis. Therefore, an absolute impression that includes not only a relative impression evaluation within a single image or only within one adjective but also an evaluation criterion for the mutual magnitude relationship between adjectives from an absolute image criterion. It becomes possible to evaluate the sensitivity. Therefore, the standardization of Qi is only standardized so that it falls within the [-1,1] interval in the average range, and sometimes it may exceed that range.

<How to determine linear combination coefficient> (see FIG. 18)
Sufficiently from among many general image model, singled out model image of a certain adjectives i, for one single energy elements E _n, Looking at the frequency distribution for their energy value, usually, common image normalization The model image is considered to have a normal distribution. However, statistically possible distribution distortion is also described within the range of the normal distribution model.

If the average value of the distribution of the model image is located in the distribution of the general image in the form of the deviation value of the model image relative to the distribution of the general image, the deviation value itself is directly Represents the importance of the energy element for that adjective. At that time, if the average value of the distribution of the general image is zero and the deviation value represented by the interval [-1, 1] at both ends is used, it can be directly used as a linear combination coefficient.
This deviation value is assumed to be a normal distribution whose width is given by the standard deviation value of the general image distribution, and is given in the form of an integral value of the distribution function of the interval from the average value of the general distribution to the average value of the model image . Further, the error of the deviation value can be calculated by giving the average value and the standard deviation value of the general image distribution and the model image distribution, respectively.

  Since the distribution of the general image may not always be a normal distribution, there are other indicators statistically defining the percentile rank using the actual distribution itself in the manner of expressing the distribution position of the model. As a result of experiments, it turned out that the deviation value given by the integrated value by the error function assuming a normal distribution is better as the linear combination coefficient than the percentile rank that integrated the real distribution from the median of the distribution. Usually you use it.

  The distribution of the energy values of the adjective model image means that the energy values fluctuate. Therefore, it can be said that the sensitivity description method that determines the sensitivity using a large number of images has a correspondence with the Gibbs distribution (canonical distribution) of statistical physics that describes the macroscopic physical quantity of the fluctuation distribution of energy values ( Reference E4).

On the other hand, if the model adjective image is reduced to the limit of only one image, the macroscopic energy of the system without the fluctuation of the energy value is fixed in the general image distribution. Replaced with a description comparing physical quantities. In that sense, a single similar image search corresponds to the macroscopic physical quantity energy value of the image, and the description of the microcanonical distribution of statistical physics where each of the momentum and angular momentum is described by a delta function. .
In that sense, the method for dynamically describing the sensibility has the capability of seamlessly describing the sensibility image search constituted by a large number of images from one similar image search. The method only changes the number of model images for determining the linear combination coefficient to the number of target image groups.

[Literature E4] Landau, Lifschitz Theoretical Physics, Vol. 5, "Statistical Physics Part 1" (3rd edition, 1976), Chapter 3, "Gibbs Distribution", Section 28, "Gibbs Distribution", and Section 35 "Gibbs distribution with variable number of particles"
<Partial integration>
Since the expected value of the adjective energy is fixed as described above for each subsystem, it is necessary to consider the integration method. In general, when the Hamiltonian describing the entire system is expressed by a linear combination of the Hamiltonians of the subsystems, that is, when the Kansei linear model is established, the wave function that is the solution of the entire Hamiltonian is It is expressed as a linear combination of the wave functions of the Hamiltonian solution. Accordingly, similarly for the eigenvalue corresponding to the wave function, the energy eigenvalue En of the entire system is represented by a linear combination of the energy eigenvalues of the subsystems. This relationship is expressed by an expression:
H = k ₁ H ₁ + k ₂ H ₂ + k ₃ H ₃ + k ₄ H ₄ + ...
H ₁ : Low-order invariant of color
H ₂ : High-order invariant of color
H ₃ : Lower order invariant of texture
H ₄ : Higher-order invariant of texture FIG. 19 is an imaginary view of the hierarchical structure of the pyramid.
The energy component of the additional factor plays a role in separating states by solving the conventional energy level degeneracy.
That is, even a slight difference in state can be quantified continuously through the expression of energy values as real values. When energy related to the composition system is added as a sub-system, the energy that has been degenerated in relation to the composition can be solved, and it becomes possible to distinguish the difference in sensitivity caused by the composition.

<How to determine linear combination coefficient k>
Subsystem adjective energy Pi, j defined here as the inner product of state density function ρ (E _n ) and linear coupling coefficient α (E _n ) for each subsystem The linear combination coefficient k can be determined through exactly the same process that has been performed for each energy element in the subsystem until now. In other words, regarding the energy index obtained by partial summation of the linear combination, it is examined how the image group of the general image model is distributed, and the deviation value indicates where the image group of the adjective model is distributed. Is used to determine the linear combination coefficient k.

  The linear combination coefficient k obtained in this way elucidates the effect ratio of the subsystem, which subsystem element of color, texture, and composition actually determines the impression of the adjective by acting strongly on the adjective. Will lead to. As a methodology, a value in the range of [-1,1] can be taken for the linear combination coefficient k of the subsystem, but when an experiment is actually performed, a result that takes only a positive value is given. This is in accord with the mystery of the role inherent in the subsystem that only changes the rate of action of the factor structure.

  In this way, when all the subsystems have the property of additive energy, the subsystems can be integrated by dividing the subsystems to determine the linear coupling coefficient again in units of partial sums. In a system that handles only energy, two-stage integration is performed that integrates with other subspaces.

In the same way, in a system that handles all of momentum, angular momentum, and energy, there is a concept of sub-system partitioning between these mechanical invariants, so it is necessary to integrate three stages. That is, after first determining the factor structure among the elements in momentum, angular momentum, and energy, which mechanical invariant of momentum, angular momentum, and energy that is the linear sum is actually the dominant role And the importance of the representative energy of the subsystems expressed by their linear sum is now determined between the subsystems.
If the process of determining the linear combination coefficient is called learning, it can be said that it is a system that performs multi-stage learning for each subsystem.

<Energy band diagram>
FIG. 20 is an energy band diagram for color and texture. If the state density function is illustrated in the order of energy levels, a scalar invariant of a lower order system becomes a discrete energy level, and a vector invariant of a higher order system becomes a continuous energy level. This is an aggregate of atoms that has an Avogadro number, for example, a metallic material, the inner electron orbit is in a discrete energy level close to the atomic orbit, and the outer electron orbit is the electron orbit of an adjacent atom. A state density diagram similar to a system with an energy band structure in a conduction band consisting of dense energy levels with overlapping layers is created.

  An example of an energy band diagram created from an adjective model “fresh” selected by a person is illustrated. For comparison, an energy band diagram in the conduction band of nickel, which is a typical transition metal, is shown (see FIG. 21).

  The electronic structure of a substance is determined by a Fermi particle system whose electron is the spin 1/2. Therefore, Pauli's exclusion principle works, and upward spins and downward spins do not mix. Therefore, the state density on the right side of FIG. 21 and the state density on the left side of FIG. 21 are not mixed, and the states are sequentially packed from the bottom of the low energy level, and the structure is filled up to the Fermi energy level.

  The energy band diagram describing the sensitivity of one image looks different. That is, a state taking a positive energy value (state density diagram on the right side of FIG. 21) and a state taking a negative energy value (state density diagram on the left side of FIG. 21) go right or left at an arbitrary level. The state density is allowed to be arbitrarily packed. This can be interpreted as describing the nature of the Bose particle in a sense. Therefore, condensation in which the state concentrates at a certain energy level may occur as a state system of the image.

  If the color describes a spin system, the Chebyshev function chosen to be optimal for describing higher-order invariants of color is actually a special function that can only describe the state of angular momentum 1 systems. It matches exactly what is there. The photon forming the image is a system having a spin angular momentum of 1 in terms of quantum mechanics. From the state of the energy band diagram, it can be seen that the color spin system describes a very detailed factor structure.

  It can be seen that the texture system described by using a spherical Bessel function suitable for describing one radial wave function describes a very rough energy structure. This is because when considering the energy level of a material atom, the radial wave function determines the large energy level, and then the wave function in the direction of the zenith angle and azimuthal angle has the next fine energy level. This is very similar to the relationship in which the spin system determines a finer energy level. However, when it comes to the electronic structure of a substance such as a solid, the spin system begins to play a major role in ferromagnets, for example, so the magnitude relationship of the degree of separation of the energy levels is agglomeration of atoms. The system cannot be described in general.

  The relationship between these energy level structures can be found by comparing the correspondence between the pyramid structure that the adjective system has and the way the pyramid structure of the predicted features is related. The nature of the general energy is very similar to the way of solving the degeneracy of the energy level of only the system considered so far by adding a new principal axis energy element.

<Image temperature>
In an image system, the temperature of the system including both steady state and unsteady state is defined. Temperature is positioned as the sum of the number of energy states. Therefore, the quantum state E _n of the energy, the temperature norm of a vector when the vector representation for state n. This corresponds to the denominator normalization factor when the energy density of states is defined. The number of energy states that each image can take is different, and the temperature is also different for each image. Therefore, the concept of image temperature can be defined. It is necessary to first define the temperature of the image in the subsystem unit that shares the accounting method of the number of states. Also, the image temperature has the same dimension as energy.

  Since the norm of the vector takes a value of zero or more, the condition that the temperature satisfies is satisfied. Also, when there is no energy state number, the image system is at absolute zero. However, since a conjugate space in which the uncertainty principle works is described as the projection plane of the subsystem at the same time, it is difficult to create an image that has absolute zero.

<Entropy of image>
The entropy of the image is defined in units of subsystems, counting the number of states based on the definition of the momentum p projected onto the subsystem and the distribution function f (p, q) in the phase space of position q. In this sense, the entropy S can be expressed as S = S (f) = S (p, q).

Since the entropy defined in the subsystem unit has an additive property, the entropy of the entire system can be defined. If a represents a subsystem,

  Entropy is a dimensionless quantity.

The entropy of subsystem a is calculated by the following formula. Integration is performed only in the interval where the value of the distribution function is finite. The interval with zero value shall be skipped. However, the number of states of the distribution function is assumed to be standardized. Therefore, it always takes a value greater than or equal to zero to satisfy the entropy requirement.

When the distribution function f (p, q) of the subsystem is projected onto a function having only momentum p, the following equation is used. The integration method is the same. However, the number of states of the distribution function is assumed to be standardized.

  A similar definition can be made when the distribution function f (p, q) of the subsystem is projected onto a function of only the position q. When the distribution functions in all subsystems are concentrated in one state, the entropy becomes zero. However, since both conjugated distribution functions that are in an uncertain relationship are viewed as distribution functions, there is no image that satisfies this condition. Entropy is a physical quantity that represents the degree of randomness of an image.

  Experiments with the entropy S to create a thermodynamic invariant of heat quantity Q = TS, for example, in a sub-system that deals only with the low-order invariant energy of color, the low temperature system is close to a monochrome image and cool The image of the impression is colorful in the high temperature system, and the image of the hot impression including the red system and the highland scenery in midsummer is divided. Also, in other subsystems, for example, in subsystems dealing with higher-order invariants of textures, images with tightly packed uniform texture are low-temperature systems, and summary images with proper composition and main subjects are displayed. It is characterized by gathering in high temperature systems. To outline other subsystems, the low-order invariant subsystems of the texture are divided into a low-temperature system with a quiet and quiet impression, and a high-temperature system that has a lot of objects and people gathered together and is full of emotion and emotion. . The high-order invariant subsystem of color is divided into a low-temperature system with a vivid impression often found in Japanese structures and sunny clothes, and a high-temperature system with a deep, dark hue. Thus, TS invariants play an important role in determining the properties of subsystems and verifying whether they are dealing with independent subsystems.

<Free energy>
The value of the energy E _n is expressed as an n-dimensional vector for the energy level n, and a scalar invariant TS is further added to define the free energy F.

This is a thermodynamic physical quantity that represents the macroscopic properties of the image system itself. From the macroscopic properties of this image, how much the properties of a certain adjective are emitted can be measured by taking the inner product with the linear combination coefficient α vector that is a model of the adjective.

Therefore, the definition with a slight change is used for the definition of the adjective energy Qi when explaining the density of states of the energy band.
1 / <S> represents the Boltzmann constant k defined in the subsystem unit. This is because the Planck's constant h for measuring the number of states in each subsystem is different, so that it plays an important role in matching both scales when integrating subsystems. Thus, when subsystem integration is performed, the denominator standardization method plays a very important role. The basic idea is that it should be brought to the denominator with the statistical average as much as possible. If the purpose is to fit the inner product to the interval [-1,1], it may be possible to bring the absolute value of the numerator to the denominator. The energy measurement as the whole system will fail.

と表され、自由エネルギーは確率分布の規格化因子の役割を果たす（文献E5参照）。そして、感性という画像系の巨視的な性質が、エネルギー関数の拘束条件E=E(p,q)のゆらぎ分布の中で、位相空間上の状態数がエントロピーS(p,q)=S(E)を通して、エネルギー状態密度分布ρ(E_n)によって記述される対応関係にあると考える。 The first reason for this change is that the shape of the inner product operation when obtaining the adjective energy is very similar to the shape of the argument of the Gibbs distribution. That is, the excitation probability for energy state E _n of the Gibbs distribution

The free energy plays the role of a normalization factor for the probability distribution (see reference E5). And the macroscopic property of the image system called Kansei is that the number of states in the phase space is entropy S (p, q) = S (in the fluctuation distribution of the energy function constraint E = E (p, q). Through E), it is considered that there is a correspondence relationship described by the energy state density distribution ρ (E _n ).

  The second reason is that free energy F = E-TS means the thermodynamic work done through them (see reference E6). This is interpreted as follows. The microscopic state of the image system (pixel value distribution) is a physical quantity representing the macroscopic state, energy E and heat quantity TS, and a thermodynamic work F is performed on the human brain. Induces sensitivity. Therefore, the energy band diagram calculated from only the signal value distribution of the image system represents the macroscopic nature of the image itself, whereas the state distribution diagram obtained by multiplying them by the linear combination coefficient α vector It is considered that the distribution map of sensibility in the brain is visually quantified.

  In fact, taking such a picture can explain the interesting facts of visual psychology. That is, it can be explained by the difference in free energy that the difference in impression between a picture with a picture frame with a white background color and a picture with a picture frame with a black background color occurs. The free energy values of the white background color and the black background color can be calculated, but the values are different. The sensibility corresponding to the amount of change in free energy when a picture is displayed in the frame from the background color is induced. Since the amount of change differs between the former and the latter, the induced sensibility is considered to be different.

  When we actually experiment with the amount of adjectives measured with free energy, we can see that the images are arranged in an extremely linear relationship. One interpretation of this is that the human visual psychological quantity has a logarithmic response characteristic with respect to the amount of light, and is given an impression of a linear scale for the argument of the excitation probability of the Gibbs distribution. it is conceivable that.

[Literature E5] Landau, Lifschitz Theoretical Physics, Vol. 5, "Statistical Physics Part 1" (3rd edition, 1976), Chapter 3, "Gibbs distribution", Section 28, "Gibbs distribution", and 31 Section "Free energy in Gibbs distribution"
[Reference E6] Landau, Lifschitz Theoretical Physics, Vol. 5 "Statistical Physics Part 1" (3rd edition, 1976), Chapter 2 "Thermodynamic quantities", Section 13 "Work and heat" , Section 15 “Free energy and thermodynamic potential”, and Section 20 “Maximum work done by objects placed in external environmental bodies”
[5] Evaluation and performance of the description model <Output of absolute adjective impression>
An example is shown in which an image is displayed and an absolute impression of an adjective model of a person is numerically evaluated based on an energy band diagram. For example, “refreshing” = + 0.47, “moist” = + 0.02, “rough” = − 0.16, “lively” =-0.75.

<Correlation matrix of adjectives>
An example of an adjective correlation matrix w _ij that clarifies the correlation between adjectives based on a certain person defined in the fifth embodiment will be shown. It is a value in the range that handles the subsystems up to color and texture. i and j are assigned adjective numbers in the order of “fresh”, “lively”, “rough”, “moist”.

<Reproducibility>
If you prepare a list that evaluates the degree of psychological impression, for example, for a population image that is a model of a general image and an image group of an adjective model selected from them, an adjective for each invariant It can also be used to determine the linear combination coefficient for constructing the model, and further, the linear combination coefficient of the subsystem that is the partial sum can be determined. If this concept is further advanced, for Qi, which is the total sum of the entire system, the average value of the distribution of Qi in the model adjective image group is up to which position as a deviation value relative to the distribution of the Qi value in the general image. It can be measured whether it was able to be extracted to the upper group. This can be an index for seeing objective recall.

  As a method of selecting a model image, there are a method of judging 0 or 1 as appropriate or not applicable, and a method of evaluating a psychological impression degree in stages. In the case of five-level evaluation, according to a method called SD method (Semantic Differential Method) in psychology, 0 is not applicable, and integers up to 1-5 are expressed with a corresponding degree of height.

  When calculating statistical data such as an average value and a deviation value, in the case of 0-1 determination, the average is calculated with an equal weight for the model image. On the other hand, in the case of five-level evaluation, when the model is averaged, the weight of 0.2 sheets applies to the psychological degree 1, the weight of 0.4 sheets corresponds to the psychological degree 2, ... As described above, the calculation is performed so that the one with a psychological degree of 5 receives the weight evaluation for one sheet.

  By calculating in this way, the error of the average value can be evaluated by evaluating the average value of the model image and the spread width of the distribution with the standard deviation. That is, in the process of calculating the deviation value indicating the position of the average value of the model image with respect to the distribution of the general image, by combining these values, the average value of the general image, and the standard deviation thereof, an error of the deviation value based on the definition is calculated. If it is also evaluated, it is possible to obtain data that combines the evaluation reliability of the deviation value.

  The reproducibility is shown as a result of selecting four adjective models of “fresh”, “lively”, “rough”, and “moist” from 254 models of general images. In the case of the method of the fifth embodiment, the statistical average result of a plurality of persons becomes the following deviation value in order, which means that an extremely high recall is realized. Deviation values of 0-100% definition are 85 ± 12%, 86 ± 7%, 98 ± 2%, 84 ± 12% respectively. Linearity can also be quantified by examining the relationship between psychological evaluation values and Qi values.

  First, in the fifth embodiment, a stable sensibility retrieval system when there are a large number of adjective model images will be described. Next, in the sixth embodiment, a sensitivity search system capable of dealing with a small number of adjective model images will be described. In the seventh embodiment, a similar image search system for a single image will be described.

[Fifth Embodiment]
(Kansei search: two-stage integration of “energy” only)
1. Conversion to Munsell HVC color space Similar to the first to fourth embodiments, both the surface with and without neutral separation are prepared as described in the fourth embodiment as the method of creating the hue surface. The separated surface is used to describe the color side, and the non-separated surface is used to describe the texture side.

  As a handling method when the hue circle is regarded as a one-dimensional axis, as described at the beginning of this embodiment, in the case of color, a point at which the frequency of the histogram is minimized is cut into one dimension. In the case of a texture, a cut is made with the Munsell value fixed at the origin. In this way, an edge surface for each color surface of the next HVC is created. This also applies to the first to fourth embodiments.

2. Creation of edge image of HVC plane The description is the same as in the second to third embodiments.
3. Creation of low-order invariants of color The symbol Fo is sometimes used for invariants as a symbol to distinguish this subsystem.
3-1. Creation of Low-Order System Distribution Function As in the first embodiment, the number of bins in the histogram is 200. The distribution function is quantized in bin units and cannot be described with higher accuracy. The distribution function is also expressed as follows.
f (H), f (V), f (C)
The variable x = H, V, C values of the distribution function f (x) are defined as Munsell values regardless of the number of bins, and the reference Munsell value is set so that the equi-step rate is satisfied between HVCs. It is assumed that it is normalized to [0,1]. However, since there is no upper limit to the value of C, the value of 1 is rarely exceeded but may be exceeded. That is,
H≡H / 100,
V≡V / 10,
C≡C / 20.

In addition, the distribution function satisfies the standardization condition. Therefore, the probability density is described.

  For f (H), the probability density of neutral hue is described in one bin called f (N).

この値をS_Foとする。 3-2. Calculation of entropy The entropy S is calculated from the distribution function f (x). When the value of the distribution function is 0, it is excluded from the integration interval in the sense that the state is not taken. When the color surface of the distribution function is distinguished by (α), entropy is calculated from the distribution functions of the H, V, and C surfaces, and the sum of those is projected to the lower-order system of the color. Represent.

This value is S _Fo .

3-3. Calculation of momentum element p _n The average value <x> and the standard deviation value σ _x are calculated from the distribution function f (x).

x=H,V,Cであるので、それぞれを運動量の要素p_nとして対応させる。
<H>, <V>, <C>, σ_H, σ_V, σ_C
色相に関連する部分は、下記のように定義する必要がある。２つに分離した成分には別の要素番号nを割り当てるものとする。

なお、運動量の要素の値は全て[0,1]の範囲で記述される。 Even for the hue plane, the average value and the deviation value are calculated on an axis obtained by making the hue circle one-dimensional using a distribution function excluding the neutral component. However, the average hue value <H> is expressed in two components as a complex number exp (2πi <H _{≠ N} >) satisfying l <H _{≠ N} > l = 1 on the hue circle with a radius of 1. Shall. At that time, the effect of excluding neutral is added to the radius representing the hue size. Therefore, the neutral ratio pop (N) is calculated. That is,

x = H, V, because in C, and to associate each as elements p _n of momentum.
<H>, <V>, <C>, σ _H , σ _V , σ _C
The part related to hue needs to be defined as follows. It is assumed that another element number n is assigned to the two separated components.

The values of the momentum elements are all described in the range [0, 1].

なお、エネルギーの要素の値は全て[-1,1]の範囲で記述される。 3-4. The calculated energy elements E _n of the energy elements E _n of is defined as follows.

The values of energy elements are all described in the range [-1,1].

部分系のエネルギーベクトルのノルムを計算すると部分系に関する画像の温度Tを定義できる。

3-5. Subsystem temperature calculation When the energy element values are collectively expressed as a vector, the energy vector of the subsystem can be defined.

By calculating the norm of the energy vector of the subsystem, the image temperature T for the subsystem can be defined.

3-6. Subsystem energy elements E _n vector of which has been calculated as the calculation of the free energy of the temperature T of the image, using the macroscopic physical quantity of entropy S, defines the free energy is a thermodynamic invariant. Free energy is a vector obtained by adding one scalar quantity to an energy vector.

Here, <> represents a statistical average regarding an arbitrary general image. Therefore, for <S>, it is necessary to calculate the entropy of an image for a large number of arbitrary general images prepared in advance and to calculate the average value thereof in advance. Physically, 1 / <S> serves as a Boltzmann constant k for linking the number of microscopic states in the phase space to a macroscopic physical quantity called entropy based on the Planck constant h defined in this subsystem. That is, the number of states in the phase space

In contrast to (s is the number of degrees of freedom of the system), the entropy is tied in the relationship of S = lnΔΓ (see Document E3). And the temperature is measured by kT via the Boltzmann constant k to connect the temperature with the energy amount. Usually, the definition is such that the Boltzmann constant is defined by including it on the entropy side, and the temperature can be described in the same scale as the energy (see Document E7). That is, S = klnΔΓ. Since the definition of the Planck constant of the quantum state of the image system varies depending on the subsystem, such an absolutely invariant physical constant cannot be defined. In view of this, an operation for standardization is performed to measure the number of states taken by each image on the basis of the average number of states that a general image can take in the sub-system. This plays the same role as the Boltzmann constant. The Boltzmann constant of this subsystem can be expressed by the following equation.

[Literature E7] Landau, Lifschitz Theoretical Physics, Vol. 5, "Statistical Physics Part 1" (3rd edition, 1976), Chapter 2, "Thermodynamic quantities", Section 9, "Temperature"
4). Creating higher-order invariants of color The symbol F is sometimes used for invariants as a symbol to distinguish this subsystem.

によって、一方でコンパクト表現されているときは他方はブロードな表現となる関係にもある。その両者の不確定性が最小となるような関数系を選ぶのが最適である。 4-0. Hilbert space representation of low-order system distribution functions
The color histogram of the HVC plane is positioned as a low-order color distribution function. The distribution function of the lower order system can be interpreted as a coordinate space q that can be measured in the original coordinate system. This is Chebyshev transformed to express the frequency and project it onto the momentum space p. This is an equivalent representation of the original distribution function from another perspective. As a basis function forming the Hilbert space, a function of a completely orthogonal system that is expressed as compactly as possible is selected in consideration of the properties of a low-order distribution function. However, the uncertainty principle of coordinate space and momentum space

Therefore, when a compact expression is given on the one hand, the other is also in a broad expression. It is best to choose a functional system that minimizes the uncertainty between the two.

(α)がHのときx=H、(α)がVのときx=V、(α)がCのときx=Cである。Nの値はN=100とする。 As in the first embodiment, it is assumed that the expansion coefficient value is defined to fall within the range [-1, 1].

When (α) is H, x = H, when (α) is V, x = V, and when (α) is C, x = C. The value of N is N = 100.

色の高次系の分布関数も2N=200のとき、kの値は200個のビンに量子化される。 4-1. Creation of higher-order distribution function The power spectrum of the Chebyshev-expanded coefficient is defined as a higher-order distribution function for color. Higher order distribution functions can be defined for the H, V, and C3 planes. Normalize to represent probability density.

When the distribution function of the higher-order color system is 2N = 200, the value of k is quantized into 200 bins.

この値をS_Fとする。 4-2. Calculation of entropy The entropy S is calculated from the distribution function f (k). When the value of the distribution function is 0, it is excluded from the integration interval in the sense that the state is not taken. When the color surface of the distribution function is distinguished by (α), the entropy is calculated from the distribution functions of the H, V, and C surfaces, and the sum of those is projected onto the higher-order system of the color. Represent.

This value to S _F.

4-3. Calculation of momentum element p _n Chebyshev expansion coefficient can be regarded as momentum in Hilbert space. Therefore, the momentum element _pn is the expansion coefficient itself.
p _n ^(α) = c _n ^(α) (α) = H, V, C.
The momentum of the different color planes will also be represented by _pn in order. These constitute the momentum p = (p ₁ , p ₂ ,..., P _i ,...) Of the phase space of this subsystem.

4-4. When building an energy matrix by a product of the calculated movement amount of elements E _n of the energy represents the kinetic energy, in the fifth embodiment by making the partial matrix in units of angular momentum, take an extended trace so as to form a complete system . Thus, the eigenvalue of energy is obtained. They become energy elements. An energy invariant is constructed basically in the same manner as in the first embodiment.

  When defining the elements of the energy, a pure spur of the energy matrix, i.e. the sum of the diagonal elements, using a relationship guaranteed by the Schwarz inequality so that each of them falls within the range [-1,1]. Standardize with. Therefore, the normalization operation cannot be performed only in the steady state constructed only by pure spurs. Standardize all other energy elements and define an extended spur. Since the value of the pure diagonal sum is defined in the range of [-1,1] for each expansion coefficient of the distribution function, the value in the range of [0,1] may be slightly extended. I want.

Since the energy eigenvalue represents the absolute amount of energy, the energy eigenvalue must have a finite value and the linear combination coefficient must have a non-zero value when finally taking these linear combinations. So it makes a different meaning. The energy eigenvalue represents the presence or absence of an energy element emitted from the image itself, and the linear combination coefficient merely represents whether or not it is an important element for a certain adjective. The energy eigenvalue must therefore be measured on an absolute basis. As for the above-mentioned problem that the diagonal sum may protrude slightly beyond the range of [0,1], the zero-point energy ε ₀ is added to the diagonal sum energy from the source, and it corresponds to the zero-point energy. You may make it define by subtracting a minute. Introducing zero energy has no effect on the linear coupling coefficient. Here, in the case of a Chebyshev expansion coefficient of a higher-order color system, a value of about 1/3 = 0.333 is introduced as the zero point energy. In the case of a higher-order spherical Bessel expansion coefficient of a texture, which will be described later, it is not necessary to introduce zero energy because the diagonal sum does not exceed the range [0, 1].

In the following, when c _n is divided into subgroups of angular momentum units, the expansion coefficient of angular momentum quantum number l = 0 is expressed as c _0n , and the expansion coefficient of angular momentum quantum number l = 1 is expressed as c _1n. To do. Therefore, the number of elements of the expansion coefficient is divided into N half. The element numbers of the subgroups shall be counted by n = 1, 2, ..., N. Only one energy element representing a pure diagonal sum is given as another definition.

Here, when a value such as c _0k = c _{0, N + n} is taken, the vector {c _0k } is _circled within a subgroup of the same angular momentum as c _{0, N + n} = c _{0, n} Connect to the ring and go back to the first point and redefine it. Similarly, c _{1, N + n} = c _{1, n} . Collectively these energy elements will be denoted by E _n.

4-5. Subsystem temperature calculation In the same way as described above, the energy vector of the subsystem and the temperature T of the subsystem can be defined.

この部分系のボルツマン定数は、この部分系のエントロピーの任意の画像の統計平均の逆数で測る。

4-6. Calculation of free energy of subsystems As before, the free energy of subsystems can be defined.

The Boltzmann constant of this subsystem is measured by the reciprocal of the statistical average of any image of the entropy of this subsystem.

5. Creating low-order invariants of textures The symbol Go is sometimes used for invariants as a symbol to distinguish this subsystem.
5-1. Creation of Low-Order System Distribution Function As in the third embodiment, the number of bins in the histogram is 200. The distribution function is quantized in bin units and cannot be described with higher accuracy. The distribution function is also expressed as follows.
f (△ H), f (△ V), f (△ C)
The value of the variable x = △ H, △ V, △ C of the distribution function f (x) is defined as a Munsell value regardless of the number of bins. -1,1]. That is,
△ H≡ △ H / 100,
△ V≡ △ V / 10,
△ C≡ △ C / 20.

In addition, the distribution function satisfies the standardization condition. Therefore, the probability density is described.

この値をS_Goとする。 5-2. Calculation of entropy The entropy S is calculated from the distribution function f (x). When the value of the distribution function is 0, it is excluded from the integration interval in the sense that the state is not taken. When the color plane of the distribution function is distinguished by (α), the entropy is calculated from the distribution functions of the edge images on the H, V, and C planes, and the sum of these is projected onto the lower-order system of the texture. Represents the entropy of.

This value is S _Go .

x=△H,△V,△Cであるので、それぞれを運動量の要素p_nとして対応させる。
<△H>, <△V>, <△C>, σ_△H, σ_△V, σ_△C
なお、運動量の要素の値は全て[-1,1]の範囲で記述される。 5-3. Calculation of momentum element p _n The average value <x> and the standard deviation value σ _x are calculated from the distribution function f (x).

x = △ H, △ V, because it is △ C, to correspond respectively as elements p _n of momentum.
<△ H>, <△ V>, <△ C>, σ _{△ H} , σ _{△ V} , σ _{△ C}
The values of the momentum elements are all described in the range [-1,1].

5-4. The calculated energy elements E _n of the energy elements E _n of is defined as follows. These values are all defined by real numbers.
(α) (α)
amam:
<△ H><△H>
<△ V><△V>
<△ C><△C>
amas:
<△ H> σ _{△ H}
<△ V> σ _{△ V}
<△ C> σ _{△ C}
asas:
σ _{△ H} σ _{△ H}
σ _{△ V} σ _{△ V}
σ _{△ C} σ _{△ C}
(β) (β):
ambm:
<△ H><△V>
<△ V><△C>
<△ C><△H>
ambs:
<△ H> σ _{△ V}
<△ V> σ _{△ C}
<△ C> σ _{△ H}
asbm:
σ _{△ H} <△ V>
σ _{△ V} <△ C>
σ _{△ C} <△ H>
asbs:
σ _{△ H} σ _{△ V}
σ _{△ V} σ _{△ C}
σ _{△ C} σ _{△ H}
The values of energy elements are all described in the range [-1,1].

5-5. Subsystem temperature calculation In the same manner as described above, the energy vector of the subsystem and the temperature T of the image of the subsystem can be defined.
5-6. Calculation of free energy of subsystems As before, the free energy of subsystems can be defined.

The Boltzmann constant of this subsystem is measured by the reciprocal of the statistical average of any image of the entropy of this subsystem.

によって、一方でコンパクト表現されているときは他方はブロードな表現となる関係にもある。その両者の不確定性が最小となるような関数系を選ぶのが最適である。 6). Creation of texture higher-order invariants The symbol G may be used for invariants as a symbol to distinguish this subsystem.
6-0. Hilbert space representation of low-order system distribution functions
The histogram of the edge image of the HVC plane is positioned as the distribution function of the texture low-order system. The distribution function of the lower order system can be interpreted as a coordinate space q that can be measured in the original coordinate system. This is expressed as a frequency by spherical Bessel transformation and projected onto the momentum space p. This is an equivalent representation of the original distribution function from another perspective. As a basis function forming the Hilbert space, a function of a completely orthogonal system that is expressed as compactly as possible is selected in consideration of the properties of a low-order distribution function. However, the uncertainty principle of coordinate space and momentum space

Therefore, when a compact expression is given on the one hand, the other is also in a broad expression. It is best to choose a functional system that minimizes the uncertainty between the two.

(α)がHのときx=△H、(α)がVのときx=△V、(α)がCのときx=△Cである。Nの値はN=100とする。 Similar to the third embodiment, in this embodiment, a description will be given of the case of expansion using s, p orbitals. In the case of expansion using s, p, d, and f orbits, the expansion can be performed in the same manner as described above.

When (α) is H, x = ΔH, when (α) is V, x = ΔV, and when (α) is C, x = ΔC. The value of N is N = 100.

6-1. Creation of higher-order system distribution function The power spectrum of the coefficient expanded in spherical Bessel is defined as a higher-order system distribution function related to texture. Higher order distribution functions can be defined for the H, V, and C3 planes. Normalize to represent probability density.

  When the distribution function of the texture higher-order system is N = 100, the value of k is quantized into 100 bins for each angular momentum. Since the value of l takes 0 and 1, it is quantized into a total of 2 × 100 = 200 bins.

この値をS_Gとする。 6-2. Calculation of entropy The entropy S is calculated from the distribution function f (l, k). When the value of the distribution function is 0, it is excluded from the integration interval in the sense that the state is not taken. When the color plane of the distribution function is represented by (α), the entropy is calculated from the distribution functions of the edge images of the H, V, and C planes, and the sum of these is projected onto the higher order system of the texture. Represents the entropy of.

This value is S _G.

6-3. Calculation of momentum element p _n The spherical Bessel expansion coefficient can be regarded as momentum in Hilbert space. Therefore, the momentum element _pn is the expansion coefficient itself.
p _n ^(α) = c _0n ^(α)
p _{N + n} ^(α) = c _1n ^(α) (α) = H, V, C.
The momentum of the different color planes will also be represented by _pn in order. These constitute the momentum p = (p ₁ , p ₂ ,..., P _i ,...) Of the phase space of this subsystem.

6-4. Calculation of energy element E _n
The expansion up to the s, p orbit is exactly the same as the formula defined for the Chebyshev expansion coefficient, which is a higher-order color system. It is only necessary to replace the Chebyshev expansion coefficient with a spherical Bessel expansion coefficient. In the case of a higher-order texture system, ε ₀ = 0 may be set for the zero point energy.

6-5. Subsystem temperature calculation In the same manner as described above, the energy vector of the subsystem and the temperature T of the image of the subsystem can be defined.
6-6. Calculation of free energy of subsystems As before, the free energy of subsystems can be defined.

The Boltzmann constant of this subsystem is measured by the reciprocal of the statistical average of any image of the entropy of this subsystem.

7). Integration of subsystem into adjective energy 7-1. Setting Adjectives Determine the adjectives to be presented in the Kansei Retrieval System using adjectives as keywords, and assign them to the i-th adjective as a symbol to distinguish them. Examples of adjectives include “fresh”, “lively”, “rough”, “moist”, and the like.

7-2. Construction of a general image model A large number of arbitrary general images are collected at random and used as a model of a general image. Usually, an image with an order of several hundreds or more is prepared. Furthermore, if it is accurate, it will be on the order of 10,000, and the more it is, the more stable the statistics will be. Using these, the distribution function p (x) of the general model image group described below is constructed. Distribution function p (x) represents the image frequency of the general image can take in unit energy interval taken by the real-valued E _n, by delimiting energy interval as the bottle, the frequency distribution function.

7-3. Construction of Adjective Model Image For each of the prepared general image models, psychological impression evaluation is performed as to whether or not the i-th adjective is applicable, and distribution data of the adjective model image is created. There are two ways to create this, and either one can be used. That is, in the case of simply corresponding or not applicable, an integer of 1 or 0 is assigned, and an image assigned 1 is used as an image having the same weight. Then, the distribution function q (x) of the adjective model image group is constructed using the corresponding image group. The other is a method in which a non-applicable case is set to 0, and the psychological absolute impression degree in the applicable case is evaluated on a 5-level scale of 1-5. In this case, in constructing the distribution data of the model, a weight of 1.0 is assigned to the 5 evaluation image, a weight of 0.8 is applied to the 4 evaluation image, and a weight of 0.6 is applied to the 3 evaluation image. A frequency distribution q (x) described below is constructed by assigning a weight of 0.4 to an image of 2 evaluations and a weight of 0.2 to an image of 1 evaluation.

7-4. One for energy elements E _n of each of the subsystems that calculated been defined above deviation in the distribution of the elements one individually, general model images takes a value of E _n distribution and E _n the adjective model image taking When the adjective model image is distributed at a position different from the average of the general model image, the energy element is considered to be working peculiar to the adjective and The weight is given in the form of a linear combination coefficient. The unique working condition can be given by the deviation value where the distribution of the adjective model image is located with respect to the distribution of the general model image.

It represents energy elements E _n in the variable x, the distribution function of the general model image p (x), the distribution function of the adjective model image and a q (x). The average value of the energy elements E _n of the general model image m _p, its standard deviation sigma _p, the average value of m _q energy elements E _n of the adjective model image, and its standard deviation sigma _q. Then, the linear combination coefficients for the energy elements E _n corresponding to the i-th adjective alpha _i (E _n) can be expressed as follows using the error function erf (x). The linear combination coefficient is defined to be 0 when positioned at the same position as the average value of the general model image, and ± 1 when positioned at both ends. Further, the fluctuation of the deviation value of the adjective model image group, that is, the error δα _i (E _n ) of the linear combination coefficient can be evaluated.

For each energy elements E _n

Deviation value and its error

であり、erf(±∞)=±1、erf(0)=0の値を返す。 here,

And erf (± ∞) = ± 1 and erf (0) = 0 are returned.

7-5. Calculation of partial energy of subsystem units To input an image and check whether it has the impression of the i-th adjective, calculate the energy element of each subsystem for the input image. Then, using the linear combination coefficient determined by the adjective model, the sum of absolute quantities of the adjective energy is obtained. In calculating the total sum, first, partial sums are taken in units of subsystems. Here, the partial sum is defined by taking the inner product of the free energy vector of the subsystem and the model adjective vector in which the linear coupling coefficient is expressed in vector. At that time, a corresponding linear combination coefficient is defined for a component obtained by adding one scalar quantity of heat quantity -TS as free energy. That is,
Free energy vector

Model adjective vector

An index for distinguishing the subsystems such as a1, a2,. a1 corresponds to a low-order system of color, a2 corresponds to a high-order system of color, a3 corresponds to a low-order system of texture, a4 corresponds to a high-order system of texture, and the like.

However, when these linear combinations are used to determine the total adjective energy Qi, the value of Pi increases as the number of energy elements that can be defined increases in the above simple inner product. Subsystems describe independent aspects, and at the time of integration of subsystems, it is necessary to treat them equally. Therefore, it is necessary to normalize so that the value of the partial sum Pi falls within the range of [-1,1]. That is,

Where <> _a represents the statistical average between all subsystems. That is, if the number of subsystems considered is n _a

Again, the standardization of the denominator follows the idea that statistical averaging should be performed as much as possible. As a result, the difference in the number of states, that is, the difference in entropy due to the difference in the number of bins of the distribution function that is defined differently depending on the subsystem is absorbed, and different subsystems can be placed on the common earth. That is, the definition of the Planck constant and the Boltzmann constant that differ depending on the subsystem is absorbed here, and the energy physical quantity Pi of a scale that can be handled in the same way is calculated. Physically, the term of the square root of the denominator can be interpreted as a correction term that is required when the physical constant varies depending on the subsystem. On the other hand, the norm average of the average temperature and adjective vector is a standardization to absorb the difference in the number of energy elements that can be defined by the subsystem. In this way, normalization based on the statistical average of the norm for placing the subsystems on the common earth plays an extremely important role in defining the absolute physical quantity.
For reference, a large amount of arbitrary general images are prepared, and the entropy value and the temperature value are experimentally examined for each sub-system. These numbers show how important the standardization procedure is to integrate subsystems.
Entropy
7.97 ≤ ~ S _Fo ≤ ~ 14.30 <S _Fo > = 12.51
2.49 ≤ ~ S _F ≤ ~ 6.05 <S _F > = 4.67
8.75 ≤ ~ S _Go ≤ ~ 14.39 <S _Go > = 12.19
7.49 ≤ ~ S _G ≤ ~ 13.82 <S _G > = 10.81
temperature
0.18 ≦ 〜T _Fo ≦ 〜1.37 <T _Fo > = 0.73
2.68 ≤ ~ T _F ≤ ~ 3.31 <T _F > = 2.91
0.003 ≤ ~ T _Go ≤ ~ 0.10 <T _Go > = 0.03
5.75 ≦ 〜T _G ≦ 〜20.63 <T _G > = 12.31
Therefore, it is a numerical value on the order of 1 / <S> to 0.1, and the square root of the denominator is a value of about 1.1 to 1.6.

8). Integration of the entire system into adjective energy Next, integrate the energy of the subsystem. In order to obtain the adjective energy Qi of the whole system, the partial sum energy Pi is linearly combined. The linear combination coefficient at this time determines whether a low-order color system, a higher-order color system, a lower-order texture system, a higher-order texture system, or the like plays an important role for an adjective i. Become. Furthermore, when composition is added as a subsystem, the importance of composition elements is taken into account.

Steps for determining the linear combination coefficient is performed by the same procedure as conducted in order to determine one single αi relative energy elements E _n described above. That is, with the partial sum Pi, j as a variable x, the average value m _p for the distribution of the general model image, its standard deviation σ _p , the average value for the adjective model image distribution m _q , its standard deviation σ _q is determined and the linear combination coefficient is determined in the same manner.

Using these linear coupling coefficients, the energy sum Qi of the entire system related to the adjective i is obtained. At this time, normalization is performed so as to be within the range of [-1,1].

Here, <> represents an average for a general model image, and <> _i represents an average for an adjective. It should be noted that the energy band diagrams shown so far are drawn using α (E _n ) in the form including the weighting coefficient k of this subsystem. In addition, the density of states is considered as α (E _n ) ρ _max (E _n ) as the maximum for the gamut region ρ _max (E _n ) that swings to the maximum in the energy elements used in the general model image. The range is painted as an energy band model.

9. Adjective search processing In this way, when an energy band model is constructed by combining a general model and an adjective model for learning, an image with an impression similar to that of the i-th adjective is used as an adjective energy for another database image. You can search based on Qi. When images are arranged in the order of Qi, it becomes a shape close to a normal distribution, and a region having a high deviation value of the upper group can be presented as a target image. Furthermore, it is also possible to present an area where the deviation value of the subgroup is low as an image of an impression having a structure opposite to this adjective. That is, the user can also know the classification result of the opposite sense image.

  Furthermore, if a single image is input and Qi is calculated for all adjectives of the adjective model prepared in advance, the degree of absolute impression received from the image for all the adjectives for that model Can be quantified. If this is rearranged and displayed in descending order, this image has a “freshness” of 0.8 and a “nodiness” as high as 0.7, and a “liveness” as low as −0.7 in the opposite direction. Therefore, a result that the impression is not felt at all can be presented.

  As an adjective energy band model, it can be an average model that is common to all people, and models specific to those cultural spheres can be constructed according to differences in countries, cultures, and languages. Alternatively, an individual energy band model that reflects differences in personal-level preferences can be constructed. Therefore, the energy band structure model adopted by this image retrieval system can be used as a tool for quantitatively elucidating the structure of human sensitivity.

Experimentally, when I tried to distinguish the difference in individuality, in contrast to the adjective “rough”, the first one is half-mountain trees and sharply forested landscape and the river flow violently splashing Kansei search results showed the effect of gathering these elements in a high rank group for the model selected to the extent. Also, while another person chose the main model image that the river flow was violently splashing, the adjective energy Qi exactly grasped the characteristics of the river flow and extracted an image close to the model Showed performance. Thus, the energy elements E _n used as feature amounts, fancy-through the entire image can be said that the capability to identify the object.

Note that the idea of obtaining the linear combination coefficient by the deviation value is extended, and the final obtained Qi value is a model once more whether or not the adjective model image is gathered in the upper group compared to the general model image. This can be confirmed by obtaining the deviation value of the image group from the general model of Qi. As a result, the validity of the theoretical model and the sufficiency of the feature amount to be handled can be verified.

Extending this idea, for the Qi for each adjective finally obtained, where is the adjective model image distribution of the i-th adjective energy Qi relative to the general model image distribution of the j-th adjective energy Qj? The j-th adjective shows how much the absolute impression of the i-th adjective is included, and the i-th adjective belongs to a similar group that gives an impression similar to the j-th adjective It is possible to obtain a correlation matrix w _ij between adjectives expressing whether they are adjectives or whether the i-th adjective belongs to a group of distant adjectives that gives the opposite impression of the j-th adjective. As a result, a map of adjective structures common to all people can be clarified, and the map structure of adjectives related to differences in cultural spheres and differences in individuality can be clarified.

By exchanging the relationship between i and j, the deviation value of the adjective model image distribution of the j-th adjective energy Qj can be similarly calculated by another path with respect to the general model distribution image of the i-th adjective energy Qi. By definition, the adjective correlation matrix is a symmetric matrix. That is,
w _ij = w _ji
However, what is calculated by following another path does not necessarily match when the characteristics of the population image are biased. Therefore, whether or not the correlation matrix of the adjective shows a value close to a symmetric matrix is to verify the validity of whether or not the image group selected in the model construction process selects a general random image. Give a good indicator. In addition, it is desirable that all diagonal components are 1 as described above. If the diagonal component has not been reached, the feature amount for capturing the adjective is insufficient, or the portion where the theoretical model hypothesis does not hold. Since it means that it exists, it gives a good evaluation index for building a search system.

[Sixth Embodiment]
(Sensitivity search: 3 steps integration of "energy + momentum + angular momentum")
In the fifth embodiment, it is described that an adjective search reflecting individuality is also possible. However, when constructing an individual adjective model, it is desirable that a certain number of images be prepared. On the other hand, when the purpose is to present a smaller number of 3 to 5 images and select from the database images that are close to their image sensitivity, the process of disappearance of information by statistical averaging is incomplete. Therefore, it is necessary to add the mechanical invariant omitted in the fifth embodiment to the object of consideration. Only the newly added process is described below.

1. 1. Conversion to Munsell HVC color space 2. Create edge image of HVC plane Creation of low-order invariants of color 3-1. Creation of low-order system distribution function 3-2. Calculation of entropy 3-3. Of defined in calculating the fifth embodiment of the elements p _n of momentum momentum, the average value <H>, <V>, the elements of the momentum only <C>.
3-4. Calculation of Angular Momentum Element _{Mn Among} the momentums defined in the fifth embodiment, standard deviation values σ _H , σ _V , and σ _C are used as angular momentum elements.
3-5. Calculation of the energy elements E _n of 3-6. Subsystem temperature calculation Follow the same procedure as described below for higher-order color systems. Details are not given here.
3-7. Calculation of free energy of subsystems Follow the same procedure as described later for higher-order colors. Details are not given here.
4). Creation of higher-order invariants of color 4-0. Hilbert space expression of low-order distribution function 4-1. Creation of high-order distribution function 4-2. Calculation of entropy 4-3. Calculation of momentum element p _n 4-4. Calculation of angular momentum element M _{n Since} the Chebyshev expansion coefficient is divided into angular momentum quantum numbers l = 0 and 1, angular momentum is calculated by taking the product of angular momentum quantum number and momentum. The angular momentum element M _n is as follows.
(c ₁₁ ^(α) + c ₁₂ ^(α) + ... + c _1N ^(α) ) (α) = H, V, C.
4-5. Calculation of the energy elements E _n of 4-6. Subsystem temperature calculation Energy element values are collectively expressed in units of energy, and momentum element values are also expressed in momentum units, and angular momentum element values are also expressed in vector units of angular momentum. Then, the energy vector, momentum vector, and angular momentum vector of the subsystem can be defined.

When the norms of the energy vector, momentum vector, and angular momentum vector of the subsystem are calculated, the image temperatures T _E , T _p , T _M related to the subsystem can be defined in units of mechanical invariants.

4-7. Calculation of free energy of subsystems From the analogy of energy, the free energy, free momentum, and free angular momentum of subsystems are defined.

5. Creation of texture low-order invariants This is exactly the same as described in “3. Creation of color low-order invariants”.
6). Create higher-order invariants for textures
When performing spherical Bessel expansion using s, p orbits, the description is omitted because it is exactly the same as the procedure described in “4.
7). Integration of mechanical invariant units of subsystems 7-1. Setting adjectives 7-2. Construction of general image model 7-3. Construction of adjective model image 7-4. Although the calculated sixth embodiment of the deviation in the distribution of the element was performed deviation values for each energy elements E _n and the calculation of the error, the elements p _n and the angular momentum that similar in the seventh embodiment This is also done for the momentum element M _n . By changing the symbols, the deviation value for the energy element is β _i , the deviation value for the momentum element is γ _i , and the deviation value for the angular momentum element is δ _i .

7-5. Calculation of Partial Energy, Partial Momentum, and Partial Angular Momentum of Mechanical Invariant Units of Subsystem As described above, model adjective vectors are determined for each of the free energy vector, free momentum vector, and free angular momentum vector. It is assumed that the deviation value is determined through the same process for the TS invariant. That is,

The inner product of both is taken and integrated into the partial energy E _i , partial momentum p _i , and partial angular momentum M _i of the mechanical invariant unit of the subsystem related to the impression of the adjective i. Normalization brings all dimensions of mechanical invariants.

8). Integration of subsystems into adjective energy The three mechanical invariants found in the subsystem are the statistically averaged expected energy value, expected momentum value, and expected angular momentum value for the adjective i in the subsystem. Represent. Of these macroscopic physical quantities, which plays an important role in the case of a few image model is represented by a coupling coefficient when the three mechanical invariants are linearly combined. Similarly, these coupling coefficients can be described by deviation values from the positional relationship between the distribution functions of the general model image and the adjective model image. Each coupling coefficient is represented by α _i (E _i ), α _i (p _i ), α _i (M _i ).

Three linear combination coefficients are represented as α _i vectors, and three mechanical invariants are represented as E ′ vectors.

The adjective energy P _i of the subsystem is expressed by taking these inner products.

9. After integration into the adjective energy of the whole system, it is exactly the same as in the fifth embodiment.
10. Adjective retrieval processing [seventh embodiment]
(Similar image search of single image: 3 steps integration of “energy + momentum + angular momentum”)
In the sixth embodiment, an example of constructing an adjective model using a minority model has been described. If you take this to the limit, you can use it to search for a single similar image. That is, it is possible to extract an image having a similar overall impression of one illustrated image through description of energy, momentum, and angular momentum, which are macroscopic physical quantities.

  The method of the sixth embodiment includes the method of the fifth embodiment. This is because as the number of model images increases, the role of momentum and angular momentum decreases, and their linear combination coefficients converge to a value close to zero automatically by statistical averaging. Therefore, the method of the sixth embodiment can be seamlessly realized from single image search to small-model sensitivity search, and further to multiple-model sensitivity search.

Claims (17)

An image input means for inputting an image;
A distribution function creating means for creating a distribution function of pixel values of the image;
A description means for describing the distribution function by expansion of the distribution function using a basis function orthogonal to each other by different weights in a complete distribution region, and describing the distribution function by an expansion coefficient;
Evaluation means for evaluating characteristics of the shape of the distribution function by calculating an index including at least a sum of quadratic forms of the expansion coefficients;
Classifying means for classifying the image into at least two categories of images having the property of the index and an image having the opposite property by rearranging the index in order of value size An image classification apparatus characterized by that. The image classification device according to claim 1,
The image classification apparatus according to claim 1, wherein the description means uses basis functions orthogonal to each other with different weights at a central portion and both end portions of a distribution region extending in a positive region. The image classification device according to claim 1,
The description means uses a basis function capable of describing the shape at a density with more emphasis than the shape of the central portion with respect to the shapes of the rising and falling portions located at both ends of the distribution function. Image classification device. The image classification device according to claim 1,
The image classification apparatus according to claim 1, wherein the description means uses a Chebyshev polynomial as the basis function. An image input means for inputting an image;
Edge image creation means for creating an edge image related to an edge component of the image;
A distribution function creating means for creating a distribution function of pixel values of the edge image;
A description means for describing the distribution function by expansion of the distribution function using a basis function orthogonal to each other by different weights in a complete distribution region, and describing the distribution function by an expansion coefficient;
Evaluation means for evaluating characteristics of the shape of the distribution function by calculating an index including at least a sum of quadratic forms of the expansion coefficients;
Classifying means for classifying the image into images having at least two categories of an image having the property of the index and an image having the opposite property by rearranging the index in order of value size. An image classification apparatus characterized by that. The image classification device according to claim 5, wherein
The image classification apparatus according to claim 1, wherein the description means uses basis functions that are orthogonal to each other with different weights near and outside the zero of the distribution range extending in the positive and negative regions. The image classification device according to claim 5, wherein
The image classification apparatus according to claim 1, wherein the description means uses a basis function capable of describing a shape at a density with more emphasis than the shape near the origin for the shape of the skirt portion located outside the distribution function. The image classification device according to claim 5, wherein
The image classification apparatus, wherein the description means uses a spherical Bessel function as the basis function. The image classification device according to claim 5, wherein
The edge image creation means filters the image to sequentially generate a high-frequency subband image having a plurality of resolutions, and sequentially integrates the high-frequency subband images from a low resolution to be integrated into one. A high frequency band image is generated, and the integrated high frequency band image is used as an edge image. In the image classification device according to claim 1 or 5,
When the image is a color image, the image is represented by three color plane images of a hue plane, a luminance plane, and a saturation plane, and the color image is obtained by performing the processing on each of the three color plane images. An image classification apparatus for classifying images into at least two categories. In the image classification device according to claim 1 or 5,
The classification means classifies the image into two adjective categories having at least opposite properties . An image input means for inputting an image;
A distribution function creating means for creating a distribution function of pixel values of the image;
Description means for describing the distribution function by expansion of the distribution function using a basis function composed of an even function group and an odd function group forming a complete system and orthogonal to each other, and describing the distribution function by an expansion coefficient;
For the function groups belonging to the same subgroup in the basis function group, the properties related to the even function and the odd function are even quantum numbers (0,2, ...) and odd quantum numbers (1,3, respectively). , ...), and a physical quantity calculating means for calculating a physical quantity obtained by summing the product of the expansion coefficient and the quantum number assigned to each of the expansion coefficients,
Evaluation means for calculating an index for evaluating characteristics of the shape of the distribution function based on the physical quantity;
Classifying means for classifying the image into images having at least two categories of an image having the property of the index and an image having the opposite property by rearranging the index in order of value size. An image classification apparatus characterized by that. The image classification device according to claim 12, wherein
When the image input means inputs an image of one color plane,
The description means uses a Chebyshev polynomial as the basis function,
The physical quantity calculating means assigns a quantum number of 0 to all even function groups and assigns a quantum number of 1 to all odd function groups to calculate the physical quantity, . The image classification device according to claim 12, wherein
When the image input means inputs an edge image related to an edge component of one color plane,
The description means uses, as the basis function, a spherical Bessel function composed of at least an even function group of degree 0 and an odd function group of degree 1;
The physical quantity calculating means assigns a quantum number equal to the order of a spherical Bessel function to each subgroup, and calculates the physical quantity. The image classification device according to claim 14, wherein
When the description means describes the distribution function using a spherical Bessel function of degree 2 or higher,
The physical quantity calculating means calculates the physical quantity, and further calculates another physical quantity based on a sum of products when the odd-order expansion coefficient is inverted to a negative sign. apparatus. The image classification device according to claim 14, wherein
The edge image is generated by filtering the image to sequentially generate a high-frequency subband image having a plurality of resolutions, sequentially integrating the high-frequency subband images from a low resolution, and combining the high-frequency subband images into one. An image classification device characterized by being a band image. The image classification device according to claim 14 or 16,
The image input device, wherein the image input means inputs an edge image of each surface of hue, luminance, and saturation.

Images