JP2004152087A - Method and apparatus for extracting feature vector of image - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To extract feature vectors reflecting variance of distribution shapes in component signal values of images, while being low in dimension for effectively performing comparison, sorting, semantic decision or the like of two-dimensional images. <P>SOLUTION: In respect to at least one kind of component signal value, an accumulated histogram of the component signal value assigned to a plurality of pixels constituting the two-dimensional images is created, and the values of n quantile points of the accumulated histogram are extracted as the feature vectors. The values of n are not smaller than 3, and natural numbers equal to or smaller than the dimensional number of the accumulated histogram. <P>COPYRIGHT: (C)2004,JPO

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a method and apparatus for extracting feature quantities representing features of a two-dimensional image, and more particularly to a method and apparatus for extracting feature quantities that have been effectively reduced in dimension.
[0002]
[Prior art]
In order to compare or classify one frame of a two-dimensional digital photographic image or moving image, it has been conventionally performed to extract a feature amount indicating the feature of the image from the image or a block image obtained by dividing them into blocks. It has been broken. Such feature amount extraction can also be used to determine what the photographing object is included in the two-dimensional image, that is, the meaning of the two-dimensional image. By determining the meaning, it is possible to improve the image quality by performing image processing only on the image region corresponding to the specific photographing target. For example, a technique has been proposed in which an image area corresponding to a person's skin is specified based on a color-related feature amount, noise is removed from only the image area, and a beautiful skin color is finished (for example, Patent Documents). 1).
[0003]
As a useful feature amount of such a two-dimensional image, one using a histogram is known. For example, for each component of the RGB color system, a histogram is created by counting the number of pixels indicating each density value from 0 to 255, and the 3 × 256 count values are used as feature amounts as they are. There is a technique in which a color space of a color system is divided into 4 × 4 × 4 areas, a histogram is created in which the number of pixels belonging to each area is counted, and each count value is used as a feature amount (for example, a patent) Reference 2).
[0004]
Also, instead of using the count value of the histogram as it is as the feature quantity of the two-dimensional image as described above, an average value and a standard deviation are obtained as representative values for each of the histograms created for hue, saturation, and brightness. A method of using it as a feature quantity is also known (see, for example, Patent Document 3). In this method, the number of extracted feature quantities, that is, the dimension of the feature quantity can be reduced as compared with the case where the count value of the histogram is used as it is as the feature quantity.
[0005]
On the other hand, although it is not in the field of image processing, especially for a frequency distribution that does not show a normal distribution, it is possible to use the value of the n quantile as a representative value of a more appropriate distribution instead of the average value or standard deviation. It has been proposed in the field of detection and the like (see, for example, Patent Document 4). Here, the “n quantile value” refers to the N / nth, 2N / nth,... N (n) from the smallest, when N values to be examined for distribution are arranged in ascending order. -1) / n points to each value.
[0006]
[Patent Document 1]
Japanese Patent Publication No. 5-62879
[0007]
[Patent Document 2]
JP 2000-353173 A
[0008]
[Patent Document 3]
JP-A-7-29007
[0009]
[Patent Document 4]
JP-A-8-166820
[0010]
[Problems to be solved by the invention]
The method of using the count value of the histogram as described above as the feature quantity of the two-dimensional image is effective, but the number of feature quantities to be extracted, that is, the dimension of the feature quantity generally increases, and the image based on the feature quantity There is a drawback that the amount of calculation when performing comparison, classification, semantic determination, etc. becomes enormous. Such a problem becomes more serious as the step size of the histogram is made finer in order to improve accuracy. On the other hand, if the step size of the histogram is increased in order to keep the feature amount dimension low, the numerical accuracy of the feature amount is lowered.
[0011]
According to the method of using the average value and standard deviation of the histogram as the feature quantity, the dimension of the feature quantity can be kept small. However, the distribution of component signal values such as the luminance component, color component, edge component, etc. that each pixel constituting the two-dimensional image shows various distribution shapes, and in particular includes a plurality of photographing objects and non-uniform textures. For images, the complexity of the distribution shape is high. Therefore, when the representative value of the histogram is used as the feature amount, it is preferable that the representative value appropriately reflects the difference in distribution shape. In this respect, the average value and the standard deviation are not necessarily appropriate as the representative values used as the feature quantity of the two-dimensional image, and may reduce the accuracy of image comparison, classification, semantic determination, etc. based on the feature quantity. Is expensive.
[0012]
In view of such circumstances, an object of the present invention is to extract, from a two-dimensional image, a feature amount that is effectively reduced in dimension while maintaining numerical accuracy and appropriately reflects a difference in distribution shape of component signal values. Is.
[0013]
[Means for Solving the Problems]
That is, the method for extracting the feature amount of the image according to the present invention is a method for extracting the feature amount representing the feature of the two-dimensional image from the two-dimensional image, which relates to at least one component signal value. Including a step of creating a cumulative histogram of the component signal values assigned to a plurality of pixels constituting a dimensional image, and a step of extracting the value of the n quantile of the cumulative histogram as a feature amount, wherein n is It is a method characterized by being a natural number of 3 or more and less than the number of dimensions of the cumulative histogram.
[0014]
An apparatus for extracting feature quantities of an image according to the present invention is an apparatus for extracting feature quantities representing features of a two-dimensional image from a two-dimensional image, and relates to at least one component signal value. Means for creating a cumulative histogram of the component signal values assigned to a plurality of pixels constituting a dimensional image, and means for extracting the value of the n quantile of the cumulative histogram as a feature quantity, wherein n is The apparatus is a natural number that is 3 or more and less than the number of dimensions of the cumulative histogram.
[0015]
Here, the “feature amount” is a general term for quantities serving as parameters indicating characteristics of a certain two-dimensional image, and includes, for example, color characteristics, luminance characteristics, depth information, edge characteristics included in the image, and the like. The indicated feature quantity may be included.
[0016]
The “component signal value” is a signal value of one component of the image assigned to each pixel of the two-dimensional image, and includes, for example, a signal value of a luminance component and a signal value of a color component. Furthermore, the signal value of each pixel of an image that has undergone some processing, such as an edge image derived from the distribution of luminance components, the absolute value of those signal values, and the standardized one, etc. "Signal value".
[0017]
The “n quantile” is one kind of representative points of the distribution. When N values to be examined for distribution are arranged in ascending order, the N / nth, 2N / Each value that is nth... N (n−1) / nth becomes a “value of n quantile”. In other words, a cumulative histogram of these N values is created, and the value of the horizontal axis corresponding to the point obtained by dividing the vertical axis into n equal parts is specified, thereby obtaining the “n quantile value”. be able to. For example, in the case of a quartile, the values on the horizontal axis corresponding to the values 0.25, 0.50, and 0.75 on the vertical axis of the standardized cumulative histogram are the values of the quartiles.
[0018]
The “dimension number” of the histogram or cumulative histogram refers to the number of columns assumed when the histogram or cumulative histogram is created.
[0019]
Further, the present invention may further determine the meaning of the two-dimensional image based on the above feature amount. Here, “determining the meaning” of a two-dimensional image refers to determining what the image is taken from. This meaning determination is preferably performed on the basis of a combination of feature amounts relating to two or more types of component signal values after extracting the feature amounts relating to the two or more types of component signal values. These two or more types of component signal values can be selected so as to include at least two of a luminance component signal value, a color component signal value, and an edge component signal value. In these cases, the above-described meaning determination may be performed using a self-organizing map.
[0020]
The value of n is preferably 20 or less.
[0021]
Further, the above-described two-dimensional image that is a target from which the feature amount is extracted may be an entire image or a block image obtained by dividing the entire image.
[0022]
Here, the “whole image” refers to a photographed digital photograph image or a two-dimensional image corresponding to the entire one frame of a moving image. On the other hand, the “block image” refers to each image piece obtained by dividing the entire image into several regions (blocks). For example, each of the entire image of 1024 × 1280 pixels divided into a size of 32 × 32 pixels. An image piece or the like corresponds to this. Note that the above “determining the meaning” includes, for example, determining which block, such as “sky”, “building”, “grass”, etc., is a captured block, This includes determining whether the whole image is a “person photograph”, a “building photograph”, a “seascape photograph”, or the like.
[0023]
【The invention's effect】
The method and apparatus for extracting the feature amount of the image of the present invention extracts the value of the n quantile of the cumulative histogram of component signal values as the feature amount, and n is a natural number less than the number of dimensions of the cumulative histogram. Therefore, it is possible to effectively compress the dimension of the feature amount and reduce the amount of calculation when performing image comparison, classification, semantic determination, and the like. In addition, since n is a natural number of 3 or more, unlike the case where the average value or standard deviation of the histogram is used as the feature amount, the difference in the distribution shape of the component signal values can be reflected in the feature amount, and the reliability can be improved. High comparison, classification, semantic determination, etc. can be performed. Furthermore, if the step size of the base cumulative histogram is made fine, even if the dimension of the feature quantity is compressed by using the value of the n quantile as the feature quantity, the numerical accuracy of the feature quantity can be kept high. it can.
[0024]
In particular, when the feature amount is used to determine the meaning of a two-dimensional image, an appropriate meaning determination can be performed even for an image that includes a plurality of shooting targets and non-uniform textures and has a complicated distribution shape of component signal values. I can expect. Furthermore, the meaning of the image is based on a combination of feature quantities related to at least two of component signal values, not just one type, for example, a luminance component signal value, a color component signal value, and an edge component signal value. If it is determined, the reliability of the semantic determination can be remarkably improved.
[0025]
In particular, in the semantic determination using the self-organizing map, it is necessary to perform a calculation for comparing the feature vector having each extracted feature quantity as a component with all of the many reference feature vectors on the self-organizing map. Therefore, the effect of dimensional compression of the above feature quantity is extremely important.
[0026]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the drawings.
[0027]
FIG. 1 is a flowchart showing the procedure of a process for specifying meaning of each image area included in a two-dimensional whole image using a method or apparatus for extracting image feature values according to the present invention. This process specifies the meaning of individual image areas in the entire image, that is, individual significant areas that are considered to correspond to any of the shooting targets such as “sky”, “building”, “grass”, etc. After that, it is useful processing for performing image processing based on conditions classified for each image area corresponding to each classification and image classification based on the meaning. First, image data representing an entire image to be processed is read in step 10, an appropriate image area is specified in step 12, the entire image is divided into block images in step 14, and features from each block image in step 16. The amount is extracted to determine the meaning of each block image, and the meaning of each image region is specified based on the extracted amount. Of these steps, the present invention particularly relates to a feature quantity extraction method and apparatus used in step 16, but the other steps will be described in order.
[0028]
An example of the image region specifying method in step 12 will be described with reference to FIG.
[0029]
FIG. 2A shows an entire image as an original image to be processed. First, with respect to each pixel constituting the original image, the color characteristics of adjacent pixels are compared and similar pixels are integrated. Here, the comparison of color characteristics and integration of similar pixels means, for example, that each component signal value of the original image represented by the RGB color system, that is, the density value of each component of R, G, and B is adjacent to each other. Each pixel is compared, and when the difference between any component signal values exceeds a predetermined threshold value, processing such as integration of those pixels is performed. Instead of the RGB color system, each component signal value represented in the YCC color system may be compared. This comparison and integration is sequentially repeated until no further integration occurs according to a predetermined criterion such as the above-described threshold value, and an area composed of pixels having similar color characteristics is expanded. The state after the integration of similar pixels is assumed to be the state shown in FIG.
[0030]
Here, among the areas constituting the image shown in FIG. 2B, an area whose peripheral length is shorter than a predetermined length is referred to as a “micro area”, and the peripheral length is equal to or greater than the predetermined length. Will be referred to as "non-micro-areas". In FIG. 2B, the areas 20 and 22 etc. are micro areas, and the areas 24, 26 and 28 etc. are non-micro areas.
[0031]
Next, each area constituting the image of FIG. 2 (b) is compared with the adjacent area, and what can be integrated is further integrated. Varies by area. For a micro area, a micro area that is completely contained in one non-micro area (eg, a micro area 20 that is completely contained in non-micro area 26) is integrated into that one non-micro area. To do. Moreover, the micro area | region which contact | connects a boundary with two or more non-micro area | regions shall be integrated into the non-micro area | region with the longer length of the boundary which touches. According to this standard, the state after the integration of the minute areas is as shown in FIG.
[0032]
For non-micro areas, the average color characteristics of the pixels forming the non-micro areas are compared with the average color characteristics of the pixels forming each adjacent non-micro area, and the degree of similarity is determined based on a threshold or the like. If there are more than non adjacent micro-areas, integration is performed. For example, for the average color feature of the non-micro-area 24 in FIG. 2c, the degree of similarity of the non-micro-area 26 with the average color feature exceeds the predetermined criteria, but the non-micro-area 28 If the degree of similarity to the average color feature is less than or equal to the predetermined criterion, the non-micro area 24 is integrated with the non-micro area 26 and is not integrated with the non-micro area 28. The final result of the integration of the non-micro area according to the predetermined criteria is, for example, as shown in FIG. Each area constituting the final image is identified as an “image area”.
[0033]
The example of the method for dividing the image area in step 12 in FIG. 1 has been described with reference to FIG. 2. Needless to say, the division into the image area in step 12 may be performed by any other known method. Yes.
[0034]
Returning to FIG. 1, in step 14, the entire image to be processed is divided into block images. In the present embodiment, it is assumed that the entire image is a digital photographic image having 1024 × 1280 pixels, and the block images are images each having 32 × 32 pixels. The divided whole image is shown in FIG. In the figure, for the convenience of explanation, it is shown in a coarser division than actual.
[0035]
Subsequently, in step 16 of FIG. 1, a feature amount is extracted from each divided block image, the meaning of each block image is determined, and the meaning of each image region is specified based on the extracted feature amount. The detailed process of the process performed in step 16 is shown in the flowchart of FIG.
[0036]
First, in step 40 of FIG. 4, a block image included in one image region among the plurality of image regions specified as shown in FIG. 2D is specified. Here, the block image included in one image region refers to a block image completely included in the image region, and does not include a block image straddling the boundary between the image regions.
[0037]
Next, in step 42, for one of the block images identified in step 40, a feature value related to luminance is extracted from the distribution of the luminance component signal values carried by each pixel included in the block image. Here, the signal value of the luminance component refers to the signal value of the Y component of the block image represented in the YCC color system in the present embodiment, but is not limited thereto.
[0038]
The processing performed in step 42 is shown in detail in the flowchart of FIG.
[0039]
First, in step 60 of FIG. 5, a histogram of signal values of luminance components carried by each pixel is created for the block image that is the current feature amount extraction target. For example, if the current block image includes only one shooting target (in this case, the sky) having substantially uniform luminance, such as the block image 30 in FIG. The shape is as shown on the left side. A block image including a plurality of photographing objects (here, a wall portion and a window portion of a building) having substantially uniform luminance, such as the block image 32 in FIG. 3, is shown on the left side of FIG. 6B. Thus, the shape has a plurality of peaks. As shown in the left side of FIG. 6C, a block image including a subject to be imaged (here, a grassland) with substantially uniform color but uneven brightness, such as the block image 34 in FIG. Become a shape. In these examples, the number of dimensions of the histogram is 256.
[0040]
Next, in step 62 in FIG. 5, a cumulative histogram indicating the cumulative count value of each signal value of the luminance component is created and normalized based on the histogram created in step 60. The cumulative histogram has the shapes as shown on the right side of FIGS. 6A, 6B, and 6C with respect to the block images 30, 32, and 34 in FIG. 3 exemplified above. As described above, the shape of the cumulative histogram is a shape unique to the block image that reflects the number of shooting targets and the texture included in the current block image.
[0041]
Subsequently, in step 64 of FIG. 5, the values of the three quartiles are specified from the normalized cumulative histogram and are used as the feature values regarding luminance. Specifically, as shown in FIG. 6, the values on the horizontal axis corresponding to the values 0.25, 0.50, and 0.75 on the vertical axis of the normalized cumulative histogram are These three values are used as feature values related to luminance. These quartile values and intervals reflect the difference in the shape of the cumulative histogram, that is, the difference in the distribution shape of the luminance component signal values in the original histogram.
[0042]
After the three feature values related to the luminance are extracted as described above, returning to FIG. 4, in step 44, regarding the same current block image, additional feature values are obtained from the signal values of the color components carried by each pixel. Extracted. In the present embodiment, as the signal values of the color components, the signal values of the two color difference components (that is, the Cr component and the Cb component) of the block image expressed in the YCC color system are used. For the signal value, a standardized cumulative histogram is created by the same method as described above, and the value of the quartile is specified as the feature amount. That is, since the values of the three quartiles are specified as the feature amounts for the two color difference components, the feature amounts relating to the six colors are extracted in total. The signal value of the color component is not limited to the signal value of the Cr component and the Cb component used in the present embodiment, and for example, the signal value of each component in the RGB color system may be used.
[0043]
Next, in step 46, an additional feature amount is extracted from the signal value of the edge component for the same current block image. In this embodiment, the signal value of the absolute value component of the vertical edge image obtained by applying a filter for edge detection as shown in FIG. 7 to the Y component image in the YCC color system of the block image, Two kinds of signal values of the absolute value component of the horizontal edge image are used as the signal value of the edge component. Again, a standardized cumulative histogram is created for the signal values of these two types of edge components by the same method as described above, and the value of the quartile is specified as a feature quantity. As a result, feature amounts relating to the features of six edges in total are extracted. Note that the signal value of the edge component is not limited to that used in the above-described embodiment. For example, the signal value of the edge image may be used as it is instead of the absolute value component, or for edge detection in four directions. You may use four types of signal values derived | led-out using the filter.
[0044]
When the extraction of a total of 15 feature values for the current block image is completed in the above steps 42 to 46, the process proceeds to step 48, and the meaning of the block image is determined. In the present embodiment, the meaning of the block image is determined using a method of mapping “feature vectors” having these 15 feature amounts as components onto a “self-organizing map”.
[0045]
As shown in FIG. 8, the “self-organizing map” is a map in which points corresponding to a plurality of reference feature vectors are two-dimensionally arranged, and points corresponding to similar reference feature vectors are close to each other. Placed in position. As shown in FIG. 8, the self-organizing map is accompanied by a map having a meaning of a corresponding size. This meaning map has a plurality of probabilities such as a probability distribution map in which the image corresponding to each reference feature vector on the self-organizing map is an “sky” image, and a probability distribution map in which the image is a “building” image. Distribution maps are superimposed. The self-organizing map and the semantic map are created in advance by causing a computer to learn feature vectors of a large number of images that are known to be “sky” and “building” images.
[0046]
An example of this learning process will be schematically described. In a state before learning, various reference feature vectors are randomly distributed on the self-organizing map. The initial value of each point in each probability distribution map is set to 0. Here, when a feature vector of one image that is known to be “empty” is input as a learning target, a reference feature that is most similar to the input feature vector is displayed on the self-organizing map. A vector is specified. This specification is performed by, for example, searching for a reference feature vector having the smallest Euclidean distance from the input feature vector. Then, the reference feature vector in the vicinity of the identified reference feature vector on the self-organizing map, for example, in the range of 7 × 7, approaches the identified reference feature vector (that is, the identified reference). Modified so that the similarity to the feature vector is increased. On the other hand, on the probability distribution map that is an image of “sky”, for example, a frequency value of “1” is added to the point corresponding to the above-described identified reference feature vector and points in the vicinity of the 7 × 7 range. Is done. Next, when a feature vector of one image known to be a “building” is input, identification of the most similar reference feature vector on the self-organizing map and a neighboring The reference feature vector is corrected. On the other hand, on the probability distribution map that is an image of “building”, a frequency value of “1” is added to a point corresponding to the identified reference feature vector and points in the vicinity thereof. When such learning is repeated, reference feature vectors indicating similar features are gradually gathered at positions close to each other on the self-organizing map. On the other hand, on each probability distribution map, an island-like frequency distribution is gradually formed. As learning progresses and similar reference feature vectors are gathered, the size of the neighborhood where the reference feature vector, which was originally 7 × 7, is corrected is gradually reduced. After completion of learning, the respective probability distribution maps are standardized and superimposed to form a meaning map as shown in FIG.
[0047]
In step 48 of FIG. 4, the feature vector extracted from the current block image is mapped to a point corresponding to the most similar reference feature vector on the self-organizing map previously derived by learning as described above. . Again, the similarity is evaluated using the Euclidean distance between both vectors as an index. Then, the point on the meaning map corresponding to the point of the mapping destination of the self-organizing map is referred to, and indicates the highest probability among the “meaning” such as “sky” and “building”. The meaning is the meaning of the current block image.
[0048]
Here, in the present embodiment, the value of the quartile is used as the feature amount, so that it is based on five different component signal values (that is, signal values of the luminance component, two color components, and two edge components). The dimension of the feature vector is suppressed to 15 dimensions, and the difference in the distribution shape of the component signal values is reflected in the feature vector. Further, by reducing the step size of the base cumulative histogram, the numerical accuracy of the feature amount can be kept high. Therefore, it is possible to perform highly reliable semantic determination while suppressing the amount of calculation required for comparing the extracted feature vector and the reference feature vector.
[0049]
Subsequently, in step 50 of FIG. 4, it is confirmed whether or not the block image included in the current image area still remains, and step 42 is performed until the meanings of all the block images included in the current image area are determined. To 50 are repeated.
[0050]
When the determination of the meanings of all the block images included in the current image area is completed, in step 52, the largest number of the determined meanings of each block image is specified as the meaning of the current image area. For example, some of the block images included in the image area 28 in FIG. 2D may be determined to mean “water”, but most are determined to mean “empty”. Therefore, the image area 28 is specified as an “empty” area.
[0051]
Subsequently, in step 54, it is confirmed whether or not there remains an image area whose meaning has not yet been specified, and the process shown in FIG. 4 is repeated until the meanings of all the image areas are specified.
[0052]
In the present embodiment, the value of the quartile is used as the feature quantity. However, the present invention is characterized in that the value of the n quartile is extracted as the feature quantity, and the value of n is 4. Any value may be used as long as it is a natural number of 3 or more and less than the number of dimensions of the cumulative histogram. When the extracted feature amount is used to determine the meaning of the image, the value of n is preferably about 3 or more and 20 or less from the viewpoint of preventing an enormous amount of calculation for the meaning determination.
[0053]
In the above embodiment, the meaning of the image is determined using the feature values extracted from the signal values of the luminance component, color component, and edge component. However, the present invention is not limited to this, and the luminance component, color component, and edge component are not limited thereto. May be used, or instead of or in addition to these, other component signal values relating to depth information or the like may be used.
[0054]
Furthermore, in the above embodiment, a method using a self-organizing map is used as a method for determining the meaning of an image. However, the present invention is not limited to this, and other methods may be used. For example, a large number of images whose meanings are known are mapped and learned in advance in the feature amount space, and the barycentric coordinates and variation (standard deviation) in the feature amount space of each meaning image are obtained. For example, a feature amount may be extracted from a certain image, mapped to the feature amount space, and the meaning may be determined based on the Mahalanobis distance from the center of gravity of each meaning image obtained by learning. Further, feature vectors that are further reduced in dimension by principal component analysis may be used.
[0055]
In the above description, as one embodiment of the present invention, the two-dimensional whole image is divided into block images, and the meaning of each block image is determined to determine the meaning of each image area included in the whole image. Although the specifying process has been described, as another embodiment of the present invention, application to the meaning determination process of the entire image is also conceivable. In this case, for example, a self-organizing map and a meaning map are created by learning in advance the entire images that are known to be “personal photographs”, “building photographs”, “seascape photographs”, etc. Then, a feature vector having a plurality of feature amounts extracted from the entire image as a semantic determination target by creating a cumulative histogram and specifying an n quantile is mapped onto a self-organizing map. Again, it goes without saying that the semantic determination method is not limited to the one using the self-organizing map.
[0056]
The image feature extraction method and apparatus according to the present invention can be applied not only to image semantic determination processing as described above but also to various other image processing such as image comparison and classification. It is.
[0057]
As mentioned above, although embodiment of this invention was described in detail, these embodiment is only an illustration, The technical scope of this invention should be defined only by the claim in this specification Needless to say.
[Brief description of the drawings]
FIG. 1 is a flowchart showing the procedure of a process for specifying meaning of each image area included in a two-dimensional whole image according to an embodiment of the present invention.
2 is a process diagram showing an example of an image area specifying method in the meaning specifying process of FIG. 1;
FIG. 3 shows an entire image divided into block images.
4 is a flowchart showing detailed steps of a meaning specifying method in the meaning specifying process of FIG. 1;
FIG. 5 is a flowchart showing in more detail the process of extracting feature values related to luminance in FIG. 4;
FIG. 6 is a diagram showing an example of a histogram of signal values of luminance components and an example of a cumulative histogram.
FIG. 7 is a diagram showing an example of an edge detection filter used for deriving edge component signal values;
FIG. 8 is a conceptual diagram showing an example of a self-organizing map and a corresponding meaning map;

Claims (16)

A method for extracting a feature amount representing a feature of a two-dimensional image from a two-dimensional image,
Creating a cumulative histogram of the component signal values assigned to a plurality of pixels forming the two-dimensional image for at least one component signal value;
Extracting the value of the n quantile of the cumulative histogram as the feature amount;
The n is a natural number of 3 or more and less than the number of dimensions of the cumulative histogram. The method according to claim 1, further comprising determining a meaning of the two-dimensional image based on the feature amount. The feature amount is extracted with respect to two or more types of component signal values, and the meaning of the two-dimensional image is determined based on a combination of the feature amounts with respect to the two or more types of component signal values. The method described. 4. The method according to claim 3, wherein the two or more types of component signal values include at least two of a luminance component signal value, a color component signal value, and an edge component signal value. The method according to any one of claims 2 to 4, wherein in the step of determining the meaning, the meaning is determined using a self-organizing map. The method according to claim 2, wherein n is 20 or less. The method according to claim 1, wherein the two-dimensional image is a whole image. The method according to claim 1, wherein the two-dimensional image is a block image obtained by dividing an entire image. An apparatus for extracting a feature amount representing a feature of a two-dimensional image from a two-dimensional image,
Means for creating a cumulative histogram of the component signal values assigned to a plurality of pixels forming the two-dimensional image for at least one component signal value;
Means for extracting the value of the n quantile of the cumulative histogram as the feature amount;
The n is a natural number of 3 or more and less than the number of dimensions of the cumulative histogram. The apparatus according to claim 9, further comprising means for determining the meaning of the two-dimensional image based on the feature amount. The feature amount is extracted with respect to two or more types of component signal values, and the meaning of the two-dimensional image is determined based on a combination of the feature amounts with respect to the two or more types of component signal values. The apparatus of claim 10. 12. The apparatus according to claim 11, wherein the two or more types of component signal values include at least two of a luminance component signal value, a color component signal value, and an edge component signal value. The apparatus according to any one of claims 10 to 12, wherein the means for determining the meaning determines the meaning using a self-organizing map. 14. The apparatus according to claim 10, wherein n is 20 or less. 15. The apparatus according to claim 9, wherein the two-dimensional image is a whole image. The apparatus according to claim 9, wherein the two-dimensional image is a block image obtained by dividing an entire image.

Images