JP5048691B2 - Image recognition method and apparatus - Google Patents

Description

  The present invention relates to an image recognition method for identifying and recognizing a specific image from input images, and in particular, specifying from an input image that has been rotated and / or scale-converted with respect to a model image including the specific image. It is related with what discriminate | determines an image efficiently.

  In the conventional image recognition method, what is called template matching is the mainstream. In this template matching, a plurality of images stored in the storage means in advance are sequentially compared with the input image (matching). The most probable image is identified to determine what the input image is.

  Among such template matching, the most common is called elastic bunch graph matching (EBGM) (see Patent Document 1), which is an algorithm devised for image recognition. In EBGM, coordinates (graphs) are defined on each of an input image and a model image, and each coordinate point on the graph is used as a matching point for discrimination.

  Specifically, at the matching points corresponding to the input image and the model image, the coordinates are set so as to maximize the evaluation function defined by the degree of similarity of the feature quantity extracted from each image and the distortion of the graph. The point is moved, and the similarity when the evaluation function is maximized is defined as the similarity between the input image and the model image, and a specific image is identified from the input image.

  As another template matching, in a calculation process for obtaining a feature value, a predetermined part in the input image (for example, if the input image is a face image, eyes, nose, nostril, mouth, chin, forehead, etc.) In particular, the size of the face image in the input image is corrected from the detected positional relationship of each predetermined part, and feature amounts extracted from a plurality of positions on the corrected face image For each of the plurality of face images stored in advance by the storage means, the similarity is calculated from the feature amount extracted from the position corresponding to the face image, and the image having the highest similarity is input. The face image is identified as a face image (see Patent Document 2).

JP-T-2004-505353 JP 2006-02049 A

  However, in the image recognition method described above, in order to increase the accuracy of recognition, feature quantities at a plurality of locations are extracted from each of one input image and model image, and moreover, The feature amount extraction uses Gabor wavelet transform, which requires a large amount of calculation. In addition, it is necessary to perform matching with all model images stored in the storage means for the input image. When performing arithmetic processing using limited hardware resources, it takes too much time and real-time processing is very difficult. In particular, the tendency is more remarkable as the number of model images increases.

  The present invention has been made to solve the above-described problems.

  The invention according to claim 1 uses a Gabor wavelet transform by a group of Gabor filters composed of a two-dimensional matrix of a plurality of frequency components and a plurality of azimuth components, from each of an input image and a model image. An image recognition method for extracting a feature amount, calculating a second feature amount from the first feature amount by a predetermined calculation, and comparing the input image with the model image, the input image and the model A first step of extracting a first feature amount using the Gabor wavelet transform with the center point of each image as a reference, and a second step using the following Equation 1 and Equation 2 from the first feature amount. A second step of calculating the feature amount of the first feature amount extracted from the input image among the second feature amount calculated in the second step. What is calculated for the plurality of azimuth components for each frequency component is an input image scale / marginalized feature amount, and the first feature amount extracted from the input image is calculated for the plurality of frequency components for each azimuth component. What is calculated is the input image rotation / marginalized feature amount, and the feature amount extracted from the model image is calculated for the plurality of azimuth components for each frequency component is the model image scale marginalized feature amount, A feature amount extracted from a model image is calculated for each of the plurality of frequency components for each azimuth component is a model image rotation / marginalized feature amount, and in each frequency component, the input image scale / marginalized feature amount and From the model image scale and marginalized feature quantity, the following number 3 is used to calculate the similarity with respect to the scale conversion parameter, and with respect to each azimuth component, the similarity with respect to the rotation conversion parameter is calculated from the input image rotation / marginalized feature quantity and the model image rotation / marginalized feature quantity by the following formula 4. And determining a scale conversion parameter having the maximum similarity among the similarities relating to the scale conversion parameter, and a rotation conversion parameter having the maximum similarity among the similarities relating to the rotation conversion parameter. An image recognition method including the fourth step.

  According to a second aspect of the present invention, in the image recognition method according to the first aspect, the maximum similarity among the similarities regarding the scale conversion parameters and the maximum similarity among the similarities regarding the rotation conversion parameters. And a fifth step of determining whether or not the specific image displayed in the model image is included in the input image.

  According to a third aspect of the present invention, first from each of an input image and a model image using a Gabor wavelet transform by a group of Gabor filters composed of a two-dimensional matrix of a plurality of spatial frequency components and a plurality of orientation components. An image recognition device that extracts a feature amount of the first feature amount, calculates a second feature amount from the first feature amount by a predetermined calculation, and compares the input image with the model image, and receives the input image An image input unit that is connected to the image input unit, stores the input image received by the image input unit, and stores the model image in advance, and the image storage unit includes: An input image storage unit for storing an input image; and a model image storage unit for storing the model image. The input image storage unit is connected to the image storage unit. An arithmetic unit that reads and calculates an image and the model image, and the arithmetic unit extracts the first feature amount using a Gabor wavelet transform based on a center point of each of the input image and the model image. A wavelet transform unit, a marginalized feature amount computing unit that calculates a second feature amount from the first feature amount using Equation 1 and Equation 2 below, and a first calculated by the marginalized feature amount computation unit Of the two feature quantities, the first feature quantity extracted from the input image is calculated for the plurality of azimuth components for each frequency component is an input image scale / marginalized feature quantity. The extracted first feature value calculated for the plurality of frequency components for each of the azimuth components is the input image processing. A marginalized feature amount, which is a feature amount extracted from the model image and calculated for the plurality of azimuth components for each frequency component, is a model image scale marginalized feature amount and is extracted from the model image A feature amount calculated for each of the plurality of frequency components for each azimuth component is a model image rotation / marginalized feature amount, and in each frequency component, the input image scale / marginalized feature amount and the model image scale / marginalized feature amount are calculated. The similarity with respect to the scale conversion parameter is calculated from the feature amount by the following equation (3), and the rotation conversion is performed by the following equation (4) from the input image rotation / marginalized feature amount and the model image rotation / marginalized feature amount in each azimuth component. Parameters A similarity calculation unit that calculates a similarity, a scale conversion parameter that is the maximum similarity among the similarities related to the scale conversion parameter, and a rotation that is the maximum similarity among the similarities related to the rotation conversion parameter The image recognition apparatus includes a conversion state calculation unit that calculates conversion parameters.

  According to a fourth aspect of the present invention, in the image recognition apparatus according to the third aspect, the maximum similarity and the rotation conversion parameter among the similarities related to the scale conversion parameter calculated by the calculation means are connected to the calculation means. Judgment means for discriminating whether the specific image displayed in the model image is included in the input image from the maximum similarity among the similarities is provided.

According to the image recognition method of the first aspect, a predetermined calculation is performed on the first feature amount extracted using the Gabor wavelet transform with reference to the center point of each of the input image and the model image, and the second feature is obtained. Calculating an input image scale / marginalized feature amount, an input image rotation / marginalized feature amount, a model image scale / marginalized feature amount, and a model image rotation / marginalized feature amount, and performing image recognition using these second feature amounts That is, the feature quantity is extracted based on only the center point of the image, and the first feature quantity obtained by the Gabor wavelet transform with a large amount of calculation is converted into a marginalized second feature quantity by a predetermined calculation. In order to perform image recognition by collecting features,
The amount of calculation by numerical calculation can be reduced as much as possible, and the numerical calculation can be performed quickly and accurately with limited hardware resources.

According to the second aspect of the present invention, in addition to the effect of the image recognition method according to the first aspect, the maximum similarity among the similarities regarding the scale conversion parameter and the maximum similarity among the similarities regarding the rotation conversion parameter. And the fifth step of determining whether or not the specific image displayed in the model image is included in the input image.
It is possible to determine whether or not the specific image displayed in the model image exists in the input image.

According to an image recognition apparatus as set forth in claim 3, the computing means for computing the data extracts the first feature amount using the Gabor wavelet transform with reference to the center point of each of the input image and the model image. A wavelet transform unit and a marginalized feature amount calculation unit that calculates the second feature amount from the first feature amount using Equations 1 and 2; The first feature value obtained by the Gabor wavelet transform with a large amount of calculation is aggregated into the marginalized feature value as the second feature value by a predetermined calculation. To do
The amount of calculation by numerical calculation can be reduced as much as possible, and the numerical calculation can be performed quickly and accurately with limited hardware resources.

According to the image recognition apparatus of the fourth aspect, in addition to the effect of the image recognition apparatus of the third aspect, the maximum similarity among the similarities regarding the scale conversion parameter and the maximum among the similarities regarding the rotation conversion parameter. Since there is a determination means for determining whether or not the specific image displayed in the model image is included in the input image from the similarity of
It is possible to determine whether or not the specific image displayed in the model image exists in the input image.

It is the schematic block diagram which showed an example of the structure of the image recognition apparatus which concerns on this invention. It is an example of an image for explaining an image recognition method concerning the present invention, and Drawing 2 (a) shows an input picture and Drawing 2 (b) shows a model picture, respectively. FIG. 3A is a schematic view showing an example of a Gabor kernel, FIG. 3A is a perspective view of the Gabor kernel, and FIG. 3B is an example of a two-dimensional view of a group of Gabor kernels. It is the figure which showed the correspondence of the input image of FIG. 2, and a model image. It is the figure which showed the relationship between a scale marginalized feature-value and a rotation marginalized feature-value while developing the frequency component and azimuth | direction component in a Gabor wavelet transform to a two-dimensional matrix. FIGS. 6A and 6B are diagrams illustrating an example of a correspondence relationship between an input image and a model image in calculation of similarity, FIG. 6A illustrates an example of a frequency component, and FIG. 6B illustrates an example of a direction component. . FIGS. 7A and 7B are diagrams illustrating experimental results of numerical experiments using the image recognition method according to the present invention, in which FIG. 7A shows the relationship between the input image size and the similarity, and FIG. 7B shows the rotation angle and the similarity. The relationship is shown respectively. It is the figure which showed the feature-value extraction point (point of a white dot) on an input image. It is the figure which showed the result of the detection rate of the numerical experiment 3.

  An embodiment of the image recognition apparatus of the present invention will be described with reference to FIG. In FIG. 1, A is an image recognition apparatus according to the present invention, and the image recognition apparatus A is a Gabor wavelet transform using a group of Gabor filters composed of a two-dimensional matrix of a plurality of spatial frequency components and a plurality of orientation components. Is used to extract a first feature value from each of the input image and the model image, calculate a second feature value from the first feature value by a predetermined calculation, and compare the input image and the model image In general, the image input unit 2, the image storage unit 3, and the calculation unit 4 are configured.

  The image input unit 2 receives, for example, an input image a (see FIG. 2A) captured by the imaging unit 1 such as a still camera or a video camera, or an input image a captured from the external storage unit 7 such as a USB memory. The input image a to be received may be either an analog signal or a digital signal. These signals are converted into monochrome digital data by a signal conversion device (not shown) provided in the image input means 2. Used by converting to

  The image storage means 3 is connected to the image input means 2 and stores the input image a received by the image input means 2 and also stores the model image b (see FIG. 2B) in advance. Known magnetic recording devices, magneto-optical recording devices, semiconductor memory devices, and the like are used. The image storage unit 3 includes an input image storage unit 31 that stores an input image a and a model image storage unit 32 that stores a model image b. Note that the model image b can also be captured from the photographing unit 1 or the external storage unit 7 via the image input unit 2 in the same manner as the input image a.

The calculation means 4 is connected to the image storage means 3, reads the input image a and the model image b from the image storage means 3, performs Gabor wavelet transform on these image data, and, as will be described later, The first feature quantity J ^L _{l, k} is obtained, or the second feature quantity X ^L _k , Y ^L _l (scale and rotation / marginalized feature quantity) is obtained from the obtained first feature quantity J ^L _{l, k} . Or calculating the correspondence S between the input image a and the model image b by obtaining the similarity Ss, Sr related to the scale and rotation, and the correspondence with respect to the scale (size) between the input image a and the model image b. In the state where the relationship (hereinafter referred to as “scale conversion state”) and / or the correspondence relationship regarding rotation (hereinafter referred to as “rotation conversion state”) is not known in advance, the most likely scale And the rotation conversion state.

As shown in FIG. 1, the computing means 4 includes a Gabor wavelet transform unit 41 that extracts a first feature quantity J ^L _{l, k} using Gabor wavelet transform, and a first feature quantity J ^L _{l. second} feature quantity X ^L _k from _k, Y ^L _l and Majinaraizudo feature calculation unit 42 for calculating a second feature quantity X _{^L k,} Y ^L _l from the scale and rotation transformation parameter s, the similarity for r The similarity calculation unit 43 for calculating Ss and Sr, and the scale and rotation conversion parameters s ′ and r having the maximum similarity Ss ′ and Sr ′ among the similarities Ss and Sr related to the scale and rotation conversion parameters s and r And a conversion state calculation unit 44 for determining '.

  By the way, the determination means 5 can be connected to the calculation means 4 described above, and this determination means 5 is the largest similarity among the similarities Ss related to the scale conversion parameter s calculated by the calculation means 4, as will be described later. A specific image (hereinafter referred to as “specific image”) displayed in the model image b in the input image a from the degree Ss ′ and the maximum similarity Sr ′ of the similarity Sr related to the rotation transformation parameter r. In this embodiment, it is determined whether or not the face image shown in FIG.

  Next, an image recognition method executed by the image recognition apparatus A configured as described above will be described below with reference to FIGS. An example in which the image input by the image input unit 2 is a human face image will be described.

  First, in the model image storage unit 32 of the image storage means 3, a model image b as shown in FIG. 2B is received from the external storage means 7 such as a USB memory as shown in FIG. It is assumed that it is taken in via the input means 2 and stored.

  Also, the face of the input image a that is the object of image recognition is captured by, for example, the image capturing unit 1 and stored in the input image storage unit 31 of the image storage unit 3 via the image input unit 2. The image a is subjected to rotation and / or scale conversion relative to the face image of the model image b, and the scale and rotation conversion state are not known in advance. It is assumed that the face image in the input image a and the model image b is located at the center of the images a and b.

Then, the data of the input image a stored in the image storage unit 3 is read into the Gabor wavelet transform unit 41 of the calculation unit 4, and the Gabor wavelet transform unit 41 uses the center point of the input image a as a reference in the first feature. The quantity J ^I _{l, k} (J ^L _{l, k} ) is extracted. The first feature amount J ^L _{l, k} is calculated by the Gabor wavelet transform expressed by the following formulas 1 and 2. The superscript “L” attached to the first feature amount J is a symbol for distinguishing the input image a from the model image b, and “I” is used for the input image a and the model image b. "M" is indicated for each of the items related to.
Here, I is an image (input image a or model image b) to be subjected to Gabor wavelet transformation, x vector is a position locus on each image, x ₀ vector is a center position coordinate of Gabor kernel, The second term in {} is a correction term for setting the direct current component to 0, and σ represents the reference size of the Gabor kernel. Further, the a vector, l, and k in Expression 1 and Expression 2 are according to Expression 3 below.
Here, a ₀ is the scale factor of the Gabor kernel, l is the frequency component, m is the number of components, and the frequency component l corresponds to the size of the Gabor kernel. Further, k is an azimuth component, and n is the number of components, which corresponds to a state where the Gabor kernel is rotated.

Similarly, the data of the model image b stored in the image storage unit 3 is also the first feature amount J ^M _{l, k} with the center point of the model image b as a reference in the Gabor wavelet transform unit 41 of the calculation unit 4. (First step S1).

  The Gabor wavelet transform (Equations 1 and 2) is an excellent conversion method that can accurately extract the features of the image from the image data. Although the amount of calculation is large, it is affected by changes in illumination. It is a difficult model and is known as a model that well represents the function of the primary visual cortex of vertebrates.

  Incidentally, FIG. 3A shows an example of the above-described Gabor kernel, which is composed of a group of Gabor kernels represented by a two-dimensional matrix of a plurality of frequency components l and a plurality of orientation components k. It is what has been. FIG. 3B is a two-dimensional illustration of this group of Gabor kernels (in this example, m = 2, n = 3).

The feature quantity J ^M _{l, k} extracted with reference to the center point of the model image b (for example, the nasal apex of the face image in FIG. 4) is the two-dimensional matrix shown in the region 10 (thick frame) in FIG. In this way, it is composed of a set of mf × n Gabor wavelet transform values. Here, mf is the number of fundamental frequency components, and n is the number of azimuth components.

Further, the feature quantity J ^I _{l, k} extracted with the center point of the input image a as a reference is (md + mf + mu) × n like the two-dimensional matrix shown in the region 11 (broken line frame) in FIG. It consists of a set of Gabor wavelet transform values. The frequency component number m of the feature quantity J ^I _{l, k} extracted from the input image a is represented by the basic frequency component number mf (shown by the solid line in FIG. 4) when represented by the polar coordinates shown on the face image of FIG. 4), the number of increments md outside the fundamental frequency component in the radial direction and the number of increments mu inside (indicated by the broken line in FIG. 4) are added.

  This is because it is unclear whether the size of the face image in the input image a is larger than the size of the face image in the model image b. This is because it is possible to appropriately cope with the size of the face image in the input image a with unknown.

  Incidentally, the scale conversion state between the input image a and the model image b is associated with the scale conversion parameter s, and the rotation conversion state between the input image a and the model image b is associated with the rotation conversion parameter r.

  Here, the number of scale conversion parameters s is (md + 1 + mu), and s = {− md, − (md−1),..., −1, 0, +1,. , Mu}, for example, when s = 0, it means that the size of the face image of the input image a is the same as that of the model image b, and when s = + 1, the input image a Means that the size of the face image is larger by one size than that of the model image b. In addition, the number of rotation conversion parameters r can be set to the same number as the number of azimuth components n. In this case, the rotation conversion parameter r has a value of r = {0, 1,..., N−1}. For example, when r = 1, it means that the input image a is rotated 180 ° / n with respect to the model image b.

More specifically, for example, a face image that is larger than the model image b shown on the right side of FIG. 4 by one scale and rotated two steps counterclockwise is input as the input image a. Then, when the correspondence relationship between the input image a and the model image b is viewed, for example, the first feature value J ^M _l, extracted from the part indicated by the point B on the model image b shown on the right side of FIG. _k is in correspondence with the first feature quantity J ^I _{l, k} extracted from the part indicated by the point A on the input image a shown on the left side of FIG.

Next, in the marginalized feature amount calculation unit 42, second feature amounts X ^L _k and Y ^L _l are obtained from the first feature amount J ^L _{l, k} obtained in the first step S1 (second step). Step S2). Here, the second feature values X ^L _k and Y ^L _k are the first feature values J ^I _{l, k} extracted from the input image a in the two-dimensional matrix shown in the region 11 of FIG. For a plurality (n) of azimuth components k for each frequency component l, the input image scale / marginalized feature value Y ^I _l calculated by the following equation 4 and the first feature value J ^I _l, extracted from the input image a _{, 5} for the plurality of ((md + mf + mu)) frequency components l for each azimuth component k and the two-dimensional input image rotation / marginalized feature amount X ^I _k calculated by the following equation 5 and the region 10 in FIG. In the matrix, the model image scale / marginalized feature obtained by calculating the first feature value J ^M _{l, k} extracted from the model image b with respect to a plurality of (n) azimuth components k for each frequency component l by the following Equation 4. and the amount Y ^M _l, model First feature amount J ^M _l extracted from Le image _b, and _k for each orientation component k for the frequency component l of a plurality (mf number), model image rotation and Majinaraizudo feature quantity X ^M calculated by the following formula 5 _{and k} . The superscript “L” added to the second feature quantities X ^L _k and Y ^L _l is a symbol for distinguishing the input image a from the model image b. "," And "M" for the model image b.

By the way, as described above, from the first feature quantity J ^L _{l, k} in the two-dimensional matrix of the input image a and the model image b, the input image scale / marginalized feature quantity Y ^I _{l as the} second feature quantity, the input image The rotation / marginalized feature quantity X ^I _k , the model image scale / marginalized feature quantity Y ^M _l , and the model image rotation / marginalized feature quantity X ^M _k are obtained from the first feature quantity J ^L _{l, This} is to reduce the number of feature quantities to be handled by replacing _k with new second feature quantities X ^L _k and Y ^L _l (each marginalized feature quantity). This is to perform image recognition quickly.

Next, the similarity calculation unit 43 calculates the following expression from the input image scale / marginalized feature amount Y ^I _l and the model image scale / marginalized feature amount Y ^M _l obtained in the second step S2 for each frequency component l. 6, the similarity Ss regarding the scale conversion parameter s is calculated, and for each azimuth component k, the input image rotation / marginalized feature amount X ^I _k obtained in the second step S2 and the model image rotation / marginalized feature amount X ^{M are obtained.} _The similarity Sr related to the rotation conversion parameter r is calculated from _k and the following mathematical expression 7 (third step S3).

Here, to give an example regarding the calculation of the similarity Ss and Sr described above, first, regarding the scale conversion parameter s, for example, in the case of the scale conversion parameter s = −1, a solid double arrow in FIG. An inner product is taken between the input image scale / marginalized feature value Y ^I _{l + s} and the model image scale / marginalized feature value Y ^M _l (numerator on the right side of Equation 6). Further, a product obtained by further normalizing this inner product (the right side of Equation 6) is the similarity Ss regarding the scale conversion parameter (s = −1). In the case of the scale conversion parameter s = + 1, the input image scale / marginalized feature value Y ^I _{l + s} and the model image scale / marginalized feature value Y ^M _l represented by the broken double arrow in FIG. The similarity Ss regarding the scale conversion parameter (s = + 1) is obtained by the same calculation.

As for the rotation conversion parameter r, the similarity Sr for the rotation conversion parameter r between the input image a and the model image b is calculated. For example, when the rotation conversion parameter r = 0, An inner product is calculated between the input image rotation / marginalized feature quantity X ^I _k and the model image rotation / marginalized feature quantity X ^M _k represented by the solid double arrow in FIG. 6B (the fractional numerator on the right side of Equation 7) ). The normalized inner product (the right side of Equation 7) is the similarity Sr related to the rotation conversion parameter (r = 0). When the rotation conversion parameter r = 1, the input image rotation / marginalized feature quantity X ^I _k is the model image rotation / marginalized feature quantity X ^{M as} represented by the broken double arrow in FIG. Similarity Sr with respect to the rotation conversion parameter (r = 1) is obtained by the same calculation with _{k + r} .

  Note that these similarities Ss and Sr indicate the degree of similarity between the input image a and the model image b. The closer the similarities Ss and Sr are to 1, the more similar the images are, and the closer to 0 they are. The closer you get, the more dissimilar the images will be.

  As described above, the similarities Ss and Sr regarding all the scale conversion parameters s and the rotation conversion parameters r are obtained. The reason why the similarities Ss and Sr are calculated for all scale conversion parameters s and rotation conversion parameters r is because the scale conversion state and rotation conversion state of the input image a with respect to the model image b are not known in advance.

Then, in the conversion state calculation unit 44, the maximum similarity Ss ′ (hereinafter referred to as “scale maximum similarity Ss ′”) among the similarities Ss related to the scale conversion parameter s obtained in the third step S3 described above. ) And the rotation conversion parameter which becomes the maximum similarity Sr ′ (hereinafter referred to as “rotation maximum similarity”) among the similarity Sr related to the rotation conversion parameter r. r ′ (see Equation 9 below) is calculated (fourth step S4), and the scale and rotation conversion state corresponding to the obtained parameters s ′ and r ′ cannot be determined in advance. It can be determined that the input image a is the most likely conversion state for the model image b. Here, the closer the scale maximum similarity Ss ′ and the rotation maximum similarity Sr ′ are to 1, the same as the specific image displayed in the model image b (the face image in this embodiment) in the input image a or It can be determined that the probability that a similar face image is included is high.

  By the way, the maximum similarity Ss ′ and Sr ′ are determined from the maximum scale similarity Ss ′ and the rotation maximum similarity Sr ′ obtained by the discriminating means 5 in the fourth step S4. It is determined whether or not the input image a includes an image that is the same as or similar to the specific image displayed in the model image b (fifth step S5). You may make it do. And these discrimination | determination results can be output by the output means 6 (refer FIG. 1), such as a display terminal.

[Verification 1]
Next, the relationship between the scale and rotation conversion state calculated by the image recognition apparatus and the image recognition method according to the present invention and the similarities Ss and Sr regarding the scale and rotation conversion parameters s and r (the scale and the rotation conversion state are correctly determined). It is verified by numerical experiment.

[Numerical Experiment 1] Relationship between Scale Conversion State and Similarity The input image a and the model image b are in the same rotation state (a state facing the same direction) and have the same size (130 × 130 ( On the basis of the state of pixel)), the similarity Ss was evaluated by increasing or decreasing the input image every four pixels in a similar manner.

[Numerical Experiment 2] Relationship between rotation conversion state and similarity The input image a and the model image b have the same size 130 × 130 (pixels), and the same rotation state (the input image a and the model image) The input image was rotated to the right or rotated to the left every 2 °, and the similarity Sr with respect to the rotation angle θ was evaluated.

The results of the numerical experiments 1 and 2 are shown in FIGS. 7A and 7B, respectively. As shown in FIG. 7A, when s = 0 (the input image and the model image have the same size), the similarity Ss regarding the scale conversion parameter s is maximized, and the maximum scale similarity Ss ′ is 1. Met. Further, as shown in FIG. 7B, when r = 0 (the input image a and the model image b have the same rotation direction), the similarity Sr regarding the rotation conversion parameter r is maximized (Sr ′). Note that the maximum rotation similarity Sr ′ did not become 1, as shown in FIG. 5, because the input image a and the model image b are calculated based on the rotation / marginalized feature values X ^I _k and X ^M _k. This is considered to be because the number m of frequency components that became different.

[Verification 2]
Next, it was verified by numerical experiments whether or not the scale and rotation conversion state between the input image a and the model image b are correctly detected by the image recognition apparatus and the image recognition method according to the present invention.

[Conditions for verification 2]
Model image: A face image of (expression D) ² = 130 × 130 pixels, no rotation (rotation angle 0.0 °), and no expression was placed in the center of an image of 256 × 256 pixels.
Input image: Five different sizes including the same size and non-rotation (130 × 130 pixels, rotation angle 0.0 °) with respect to the face image of the model image b at the center of the 256 × 256 pixel image Transformation (83 × 83, 104 × 104, 130 × 130, 163 × 163, 204 × 204 pixels) and 8 different rotational transformations (rotation angles 0.0 °, 22.5 °, 45.0 °, 67. A total of 40 different scale and rotation-transformed face images having 5 °, 90.0 °, 112.5 °, 135.0 °, and 157.5 °) were arranged.
The verification was performed on 100 different face images.

[Verification Method for Verification 2]
(1) Determination of whether or not the scale and rotation conversion state can be detected correctly The scale parameter s ′ taking the maximum scale similarity Ss ′ is the same as the scale conversion state actually applied to the input image, and the rotation maximum When the scale parameter r ′ taking the similarity Sr ′ is the same as the rotation conversion state actually applied to the input image, it was determined that the scale and rotation conversion state were correctly detected.
(2) Judgment of effectiveness of this image recognition method The detection rate was calculated | required by the following formula, and it determined based on the numerical experiments 3-6 which changed various conditions.
Detection rate (%) = (number of images in which both scale and rotation conversion state are correctly detected) / (number of test images = 40 sheets / person × 100 people) × 100

[Numerical experiment 3] Dependence due to positional deviation A point shifted by d (pixel) (relative distance Sp = d / D) from the center point with respect to the size D (pixel) of the face image (white color in FIG. 8) The detection rate was obtained by extracting the feature amount based on the dot).

[Numerical experiment 4] ... Dependence on facial expression Using an input image a that is non-rotating, has the same size, and has a different facial expression with respect to the model image b, extracts the feature amount based on the center point of each image, and determines the detection rate. Asked.

[Numerical experiment 5] ... Dependence by facial expression, positional deviation, scale / rotation Combining numerical experiments 3 and 4, the model image b is converted to scale and rotation, and the facial expression is different. The detection rate was obtained by extracting the feature value based on the point shifted from the center point.

[Numerical Experiment 6]... Dependence by Person In Numerical Experiment 3, the input image a and the model image b were made different from each other to extract feature amounts, and the detection rate was obtained.

  The results of the detection rates of the numerical experiments 3 to 6 are shown in FIG. Regarding the dependency due to the positional deviation in the numerical experiment 3, as shown in FIG. 9, when the relative distance Sp (= d / D) from the center point is about 0.1 or less, the detection rate is 98% or more, The result that almost correct detection was carried out was obtained (the detection rate of the features extracted from the center point was 100% in all cases). Further, as can be seen from the difference between the detection rates of the numerical experiments 3 to 5 and the detection rate of the numerical experiment 6 shown in Table 1, the position, scale and rotation of the feature amount extraction on the input image a, facial expression It was possible to identify the same person with a high probability even when changing.

  By the way, in the numerical experiment mentioned above, it showed about the verification result using a human face image, but it is an experiment similar to the numerical experiments 3-6, Comprising: As a model image b, 21 types of dog images, Similar results were obtained with 19 types of cat images and 20 types of car images.

A image recognition device 2 image input means 3 image storage means 4 calculation means 5 discrimination means

Claims (4)

A first feature amount is extracted from each of the input image and the model image by using Gabor wavelet transform by a group of Gabor filters composed of a two-dimensional matrix of a plurality of frequency components and a plurality of orientation components. An image recognition method for calculating a second feature value from a feature value of one by a predetermined calculation and comparing the input image with the model image,
A first step of extracting a first feature amount using the Gabor wavelet transform with reference to a center point of each of the input image and the model image;
From the first feature amount, a second step of calculating a second feature amount using the following Equation 1 and Equation 2 below, and among the second feature amount calculated in the second step,
What is calculated for the plurality of azimuth components for each frequency component of the first feature amount extracted from the input image is an input image scale marginalized feature amount,
What is calculated for the plurality of frequency components for each azimuth component of the first feature amount extracted from the input image is the input image rotation and marginalized feature amount,
What is calculated for the plurality of azimuth components for each frequency component the feature amount extracted from the model image is a model image scale marginalized feature amount,
A feature amount extracted from the model image is calculated for the plurality of frequency components for each azimuth component is a model image rotation / marginalized feature amount,
In each of the frequency components, a similarity with respect to a scale conversion parameter is calculated from the input image scale / marginalized feature quantity and the model image scale / marginalized feature quantity by Equation 3 below, and in each azimuth component, the input image rotation A third step of calculating a degree of similarity related to the rotation transformation parameter from the marginalized feature quantity and the model image rotation / marginalized feature quantity by the following equation 4;
A fourth step of determining a scale conversion parameter having the maximum similarity among the similarities regarding the scale conversion parameter and a rotation conversion parameter having the maximum similarity among the similarities regarding the rotation conversion parameter; An image recognition method characterized by
  Furthermore, the specific image displayed in the model image is included in the input image based on the maximum similarity among the similarities regarding the scale conversion parameters and the maximum similarity among the similarities regarding the rotation conversion parameters. The image recognition method according to claim 1, further comprising a fifth step of determining whether or not the image is present. The first feature quantity is extracted from each of the input image and the model image by using the Gabor wavelet transform by a group of Gabor filters composed of a two-dimensional matrix of a plurality of spatial frequency components and a plurality of orientation components. An image recognition device that calculates a second feature value from a first feature value by a predetermined calculation and compares the input image with the model image,
Image input means for receiving the input image;
An image storage means connected to the image input means for storing the input image received by the image input means and storing the model image in advance, and the image storage means for inputting the input image An image storage unit, and a model image storage unit for storing the model image,
An arithmetic unit connected to the image storage unit and reading and calculating the input image and the model image from the image storage unit;
A Gabor wavelet transform unit for extracting the first feature amount using a Gabor wavelet transform based on a center point of each of the input image and the model image;
A marginalized feature amount computing unit that calculates a second feature amount from the first feature amount using the following Equation 1 and Equation 2;
Of the second feature amount calculated by the marginalized feature amount calculation unit,
What is calculated for the plurality of azimuth components for each frequency component of the first feature amount extracted from the input image is an input image scale marginalized feature amount,
What is calculated for the plurality of frequency components for each azimuth component of the first feature amount extracted from the input image is the input image rotation and marginalized feature amount,
What is calculated for the plurality of azimuth components for each frequency component the feature amount extracted from the model image is a model image scale marginalized feature amount,
A feature amount extracted from the model image is calculated for the plurality of frequency components for each azimuth component is a model image rotation / marginalized feature amount,
In each of the frequency components, a similarity with respect to a scale conversion parameter is calculated from the input image scale / marginalized feature quantity and the model image scale / marginalized feature quantity by Equation 3 below, and in each azimuth component, the input image rotation A similarity calculation unit that calculates a similarity with respect to a rotation conversion parameter from the marginalized feature quantity and the model image rotation / marginalized feature quantity by the following Equation 4.
A conversion state calculation unit that calculates a scale conversion parameter that is the maximum similarity among the similarities regarding the scale conversion parameter and a rotation conversion parameter that is the maximum similarity among the similarities regarding the rotation conversion parameter; An image recognition apparatus characterized by comprising:
  A model is connected to the calculation means from the maximum similarity of the scale conversion parameters calculated by the calculation means and the maximum similarity of the rotation conversion parameters. 4. The image recognition apparatus according to claim 3, further comprising a discriminating unit that discriminates whether or not the specific image displayed in the image is included.