KR101521136B1

KR101521136B1 - Method of recognizing face and face recognition apparatus

Info

Publication number: KR101521136B1
Application number: KR1020130156639A
Authority: KR
Inventors: 박혜영; 서정인
Original assignee: 경북대학교 산학협력단
Priority date: 2013-12-16
Filing date: 2013-12-16
Publication date: 2015-05-20

Abstract

Disclosed is a method for recognizing a face including the steps of: receiving a test image, dividing the test image into a plurality of blocks, and then obtaining a plurality of local feature descriptors for each divided block; selecting one local feature descriptor for each block of the test image by using an object probability model; calculating a weighted value for each block of the test image; calculating distance values of similarity between the blocks, matched with each other, by using the selected local feature descriptor and the weighted value while sequentially comparing a plurality of reference images with the test image; and recognizing an image among the reference images which corresponds to the test image by comparing the calculated distance values of similarity.

Description

[0001] METHOD OF RECOGNIZING FACE AND FACE RECOGNITION APPARATUS [0002]

The present invention relates to a face recognition method and a face recognition apparatus, and more particularly, to a face recognition method and a face recognition apparatus that exhibit robust recognition performance for face images obtained in various environments.

Awareness has been a subject of long research in the field of computer vision and machine learning in recent decades. In spite of the great interest in image recognition, conventional image recognition researches have proved their research results through experiments using images acquired in a limited environment.

Particularly, in face recognition, there are deformations such as occlusion, movement, and rotation as well as illumination and facial expression in images acquired in a general environment. This has been a factor that significantly degrades the performance of face authentication system using face image.

Therefore, there is a need for a device capable of maintaining a robust image recognition performance even for images having various distortions detected in a real-time environment.

It is an object of the present invention to provide an image recognition method and an image recognition apparatus that exhibit robust recognition performance even for images obtained in various environments.

According to an aspect of the present invention, there is provided an image recognition method for dividing a test image into a plurality of blocks and obtaining a plurality of regional feature descriptors for each of the divided blocks, Selecting one local feature descriptor for each block of the test image using an object probability model estimated based on a plurality of reference images, using the object probability model, Calculating the similarity distance values between blocks matching each other using the selected local feature descriptor and the weight value while sequentially comparing the plurality of reference images and the test image, Comparing the similarity distance values, and comparing the similarity distance values to correspond to the test image among the plurality of reference images Includes a step of recognizing an image.

In this case, the image recognition method according to an embodiment of the present invention includes the steps of: dividing a plurality of reference images into a plurality of blocks and obtaining one regional feature descriptor for each block, using the obtained local feature descriptor And estimating and storing the object probability model.

The obtained plurality of local feature descriptors have different directional properties, and the selecting may include, for each block of the test image, one of the obtained plurality of local feature descriptors corresponding to the directionality of the test image Can be selected.

In this case,

, Where m is the index of each block, < RTI ID = 0.0 >

(K = 0, ..., 8) of the m block of the test image, p _m is the object probability model,

(M = 1, ..., M) of m blocks of the test image, and I ^tst may be a test image.

Meanwhile, in the image recognition method according to an embodiment of the present invention, a determination is made as to whether to reject the recognition of the test image using the one local feature descriptor selected per each block, the weight per each block, Wherein the calculating of the similarity distance values and the recognition of the image may be selectively performed according to the determination result of the determination step.

In this case, if the value calculated by applying the selected local feature descriptor to the object probability model is less than the first threshold value, the determining step may reject the recognition of the test image.

If the similarity distance value between the selected image and the test image is greater than the second threshold value, the test image is displayed on the test image, You can refuse recognition.

The calculating of the similarity distance values may include:

And calculates the similarity distance values according to the following equation: < RTI ID = 0.0 >

Is the local feature descriptor of m blocks of the ith reference image, d () is the similarity distance calculation function,

Is a weight for the m block of the test image, and I ⁱ may be the i-th reference image.

Meanwhile, the image recognition apparatus according to an embodiment of the present invention includes: a storage unit for storing an object probability model estimated using a local feature descriptor for a plurality of blocks constituting each of a plurality of reference images; An input unit for dividing the test image into a plurality of blocks, acquiring a plurality of local feature descriptors for each divided block, and detecting one local feature descriptor for each block of the test image using the object probability model A calculating unit for calculating a weight per block of the test image using the object probability model, and a calculating unit for sequentially comparing the plurality of reference images and the test image, using the selected local feature descriptor and the weight value Calculates similarity distance values between blocks that match each other, Comparing the Li values, and includes an image recognition from among the plurality of reference image recognition the image corresponding to the test image.

In this case, the obtained plurality of local feature descriptors have different directional attributes, and the detecting unit may detect, for each block of the test image, one of the obtained plurality of local feature descriptors, A local feature descriptor can be detected.

In this case,

, Where m is the index of each block, < RTI ID = 0.0 >

(M = 1, ..., M) of m blocks of the test image, and I ^tst may be a test image.

The image recognition apparatus according to an embodiment of the present invention may further include a determination unit for determining whether to reject the recognition of the test image, and the image recognition unit may selectively operate according to a determination result of the determination unit have.

In this case, the determination unit may reject the recognition of the test image if the value calculated by applying the detected local feature descriptor to the object probability model is less than the first threshold value.

If the similarity distance value between the selected image and the test image is greater than the second threshold value, the determination unit may select the image having the smallest similarity distance value from the test image among the plurality of reference images, I can refuse.

Meanwhile,

According to the above-described various embodiments, image recognition can be performed considering various variables included in the input image, so that excellent recognition performance can be expected.

1 is a diagram for explaining a method of estimating an object probability model according to an embodiment of the present invention;
FIG. 2 and FIG. 4 are flowcharts for explaining an image recognition method according to various embodiments of the present invention;
3 is a diagram for explaining feature selection according to an exemplary embodiment of the present invention;
FIG. 5 is a block diagram schematically showing an image recognition apparatus according to an embodiment of the present invention, and FIG.
FIG. 6 through FIG. 8 show various experimental examples to which the present invention is applied.

Hereinafter, an image recognition method and an image recognition apparatus according to the present invention will be described in detail with reference to the accompanying drawings.

According to the present image recognition method, an image most similar to a test image among a plurality of reference images is detected using the object probability model generated based on a plurality of reference images, despite the movement of the test image direction, can do.

The plurality of reference images are images in which an arbitrary object is photographed, which means an image to be learned for estimating an object probability model. The test image means an object to be recognized.

The image recognition method can be divided into a learning step and a recognition step. In the learning step, a plurality of reference images are learned to estimate the object probability model, and in the recognition step, feature selection for the test image is performed. First, a method of estimating an object probability model in a learning step will be described with reference to FIG.

First, an object image 1, which is one arbitrary reference image as shown in FIG. 1, is divided into a plurality of blocks (M). For example, when the recognition target object is a face, the image recognition according to an embodiment of the present invention uses the entire face image. Therefore, all the parts other than the specific part (eye, nose, mouth, etc.) of the object image 1 are partitioned into blocks of the same size. At this time, the block may be divided into a grid pattern. Each of the divided blocks is hereinafter referred to as P1 to PM.

Next, one local descriptor is obtained for each of the blocks P1 to PM divided into a plurality of blocks. For example, local feature descriptors can be obtained through SIFT, HOG, and LBP. Here, SIFT, HOG and LBP are well known in the art, and a detailed description thereof will be omitted.

A case in which a Scalar Invariant Feature Transform (SIFT) is used to acquire a regional feature descriptor will be described as an example. The process of acquiring the regional feature descriptor is as follows.

SIFT goes through two computation steps to create a set of image features of the object image (1). The first step determines how to select the important pixels from the reference image (1). Here, the selected important pixel is referred to as a 'keypoint'. Step 2 defines an appropriate descriptor for the selected keypoints to represent meaningful local properties of the reference image (1). Here, the descriptor is called a 'local feature descriptor'. Thus, one object image 1 can be represented by a regional feature descriptor of each of the blocks P1 to PM.

SIFT uses a scale-space Difference-Of-Gaussian (DOG) function to detect multiple keypoints within each image. For an input object image I (x, y), the scale-space is a function L (x, y, y) provided from a convolution of a variable-scale Gaussian G (x, y, σ). Accordingly, the DOG function is defined as Equation (1).

[Equation 1]

Here, x is an x-axis coordinate, y is a y-axis coordinate, σ is a scale, and k is a multiplicative factor. In this case, the local maximum and minimum values of the D (x, y, sigma) function are based on eight neighboring blocks surrounding a block in the object image (1).

On the other hand, in case of face recognition, a very small number of keypoints can be extracted because of lack of texture of face image. In this case, Dense-SIFT extraction method, which is another local feature extraction method, is applied instead of SIFT The above problem can be solved.

Each local feature descriptor extracted through the SIFT extraction method is represented by a 128-dimensional vector descriptor, which is a four-part 128-dimensional vector. Where the four parts are locus, scale (sigma), direction and slope that indicate where the feature is selected. In this case, the degree of inclination m (x, y) and the direction (? (X, y)) of each keypoint located in the two-dimensional coordinate (x, y) Lt; / RTI >

&Quot; (2) "

&Quot; (3) "

Then, in order to apply the SIFT extraction method to represent the object image 1, the location of the regional feature descriptors is first determined on the M local feature descriptors and on a regular grid. Here, since each local feature descriptor is represented by a descriptor vector y _m , one face image I can be represented by a set of M descriptor vectors. In this case, the face image I can be expressed by the following equation (4).

&Quot; (4) "

If a plurality of reference images are {I ⁱ } _i _{= 1, ... ,} _N , the entire plurality of reference image sets T can be expressed by the following equation (5).

&Quot; (5) "

Then, an 'object probability model' can be derived using the reference image set T. For example, when the reference image is a face image, a probability model of a normal face can be derived. In order to obtain the object probability model, first, the reference image set T is divided into M subsets according to the positions of keypoints as shown in Equation (6) below. Here, m is the index of each block.

&Quot; (6) "

Using each subset T _m , the object probability model (p _m ) for the regional feature corresponding to the m block can be estimated using simple Gaussian estimation as shown in Equation (7) below.

&Quot; (7) "

In Equation 7, Z is a normalization factor. And the two model parameters as a mean (μ _m) and covariance matrix (Σ _m) can be estimated by a simple average and the simple covariance matrix of each subset (T _m).

Using the estimated object probability model as described above, the probability that each local feature descriptor is found at a specific position of the input test image can be calculated.

Hereinafter, a step of recognizing a test image using the object probability model will be described with reference to the flowchart of FIG.

Referring to FIG. 2, first, a test image is input (S210). The test image is the image to be compared with the reference image.

The input test image is divided into a plurality of blocks so as to correspond to the blocks of the reference image (S220). For example, if the reference image is divided into M blocks as described above, the test image is also divided into M identical segments.

A plurality of local feature descriptors are obtained for each block of the divided test image (S230). Since a plurality of regional features are extracted for each block, more features are extracted than those extracted from the reference image. Specifically, in the test image, more features are extracted by reducing the keypoint interval than the reference image. Therefore, there is no overlap between each block of the reference image, but there is a 50% overlap between adjacent two partial areas in the test image. For example, as shown in Fig. 3, in the comparison between the general face model (a) and the test image (b) estimated from a plurality of reference images, in the m blocks corresponding to each other, Can see more. A plurality of local feature descriptors of m blocks of the test image (

) Can be expressed by the following equation (8).

&Quot; (8) "

Next, one local feature descriptor is selected for each block of the test image (S240). That is, a plurality of local feature descriptors (

) Candidate (c) is a local feature descriptor representing m blocks

). According to one embodiment of the present invention, a plurality of local feature descriptor candidates c of the block m have different directional properties. For example, as shown in Fig. 3, a position corresponding to a key point of the general face model (a) Geographic feature descriptor

, And eight local feature descriptors ("

, ... ,,

) Have different directional properties. Multiple local feature descriptor candidates are applied to the object probability model, and when this process is repeated for all blocks, k _max , from which the highest probability value is obtained, can be obtained. This is obtained through the following equation (9).

&Quot; (9) "

here,

(K = 0, ..., 8) of the m block of the test image.

For example, if the local feature descriptor selected in FIG. 3 through equation (9) is eta ₅ , this would mean that the test image (b) has the following directionality. Actually, the test image (b) of FIG. 3 shows that when the lower lip is cut off, its directionality is lower when compared with the general face model (a).

According to the above-described embodiment, it is possible to know in which direction the moving image is deformed based on the general face image of the test image, and the face recognition performance for the moving image can be improved by selecting the characteristic of the direction.

It is possible to acquire the local feature descriptor of the test image I ^tst as shown in Equation (10) using k _max obtained through Equation (9). For example, when? ₅ is selected, since the regional feature descriptors having a direction corresponding to? ₅ for each block of the test image are selected for each block of the test image, . In this case, k _max becomes 5.

&Quot; (10) "

here

(M = 1, ..., M) of m blocks of the test image.

In the next step, a weight for each block of the test image is calculated using the object probability model (S250). Specifically, the local feature descriptors for the test image obtained are applied to the object probability model as shown in Equation (10), and the weight of each block of the test image is calculated. Here, the weight w _m can be expressed by Equation (11).

&Quot; (11) "

For example, if there is an occlusion such as occluded in a test image, the weight for the corresponding block will be measured low.

Then, the obtained test image I ^tst And a plurality of reference images ^Ii are successively compared with each other to calculate inter-block similarity distance values matching each other (S270). In this case, the lower the value of the similarity distance value, the higher the degree of similarity between the reference image and the test image. The similarity distance value can be calculated using the following equation (12).

&Quot; (12) "

here

Is a regional feature descriptor of m blocks of the i-th reference image, d (·, ·) is a similarity distance calculation function,

Is a weight for m blocks of the test image, and I ⁱ is an i-th reference image.

Finally, the calculated similarity distance values are compared, and an image corresponding to the test image among the plurality of reference images, that is, the image most similar to the test image is recognized (S270).

4 is a flowchart illustrating an image recognition method according to another embodiment of the present invention. Hereinafter, a description will be given of a method for enhancing the reliability of image recognition by previously rejecting a test image that is too similar to the reference image. The embodiment described below further includes the image rejection step in the above-described embodiment.

Referring to FIG. 4, the image recognition method according to the present embodiment can be largely divided into a learning process and a recognition process. In FIG. 4, the solid line represents the learning process, and the dotted line represents the recognition process.

First, a plurality of reference images are stored (S410). Then, each of the plurality of reference images is divided into a plurality of blocks, and a regional feature descriptor is obtained for each divided block (S412). Then, the object probability model is estimated using the obtained local feature descriptor (S414). Then, a test image is input (S420), and a plurality of regional feature descriptors for each of the divided blocks of the test image are obtained as described above (S422).

Then, in consideration of the directionality of the test image, a part of the obtained local feature descriptors is selected (S424), and is defined as a regional feature descriptor indicating a test image (S424). The above steps have not been described in detail.

In the next step, it is determined whether to reject the test image using the estimated object probability model (S426). If the test image is rejected at this stage, image recognition is no longer performed. According to one embodiment, the rejection of an image may be determined through a determination according to Equation (13) below. This step is called the first image rejection.

&Quot; (13) "

if

, The test image I ^tst is rejected.

Here,? Is a first threshold value arbitrarily set.

In addition, the image rejection method according to another embodiment may be additionally or independently considered in the image rejection method described above. This is called a second image rejection. Specifically, if the image is rejected, an image having the smallest similarity distance value is selected from among the plurality of reference images. If the similarity distance value between the selected image and the test image is greater than the second threshold value, the test image is rejected. For example, a plurality of reference images are classified using a K-nearest neighbor classifier, and the nearest reference image (I ^nrst ) is detected. The detection of I ^nrst uses the following equation (14).

&Quot; (14) "

If the similarity distance value to which the detected I ^nrst is applied is larger than the arbitrarily set second threshold value, the test image is rejected.

If the image is not rejected, the similarity distance value between the plurality of reference images and the test image is calculated (S430). Meanwhile, since the above-described second image rejection step is also performed through the calculation of the similarity distance value, it can be simultaneously performed in the similarity distance value calculation step S430 of FIG.

In the next step, the calculated values are compared with each other to finally recognize the image (S432).

Using the image rejection step as described above, there is an advantage that an object detector, for example, a detector for detecting a face, can effectively process images in which a portion other than a face is erroneously detected.

The above-described image recognition method can be implemented in an image recognition apparatus. 5 is a block diagram schematically showing the configuration of an image recognition apparatus according to an embodiment of the present invention.

Referring to FIG. 5, the image recognition apparatus 100 includes a storage unit 110, an input unit 120, and a control unit 130. Meanwhile, the image recognition apparatus 100 may be implemented by various electronic apparatuses such as a smart phone, a tablet PC, a desktop PC, and the like.

The storage unit 110 stores various programs and data used in the image recognition apparatus 100. In particular, the storage unit 110 may store a plurality of reference images to be learned to generate an object probability model. The reference image may be stored in the image recognition apparatus 100 or may be transmitted from an external apparatus through the communication unit 140. [ The storage unit 110 also stores an object probability model estimated through learning of a plurality of reference images.

The input unit 120 is configured to input a test image to the image recognition apparatus 100. For example, the input unit 120 may be a photographing apparatus that photographs an object, a configuration that receives a test image as an interface, and the like.

The control unit 130 can control the overall operation of the image recognition apparatus 100. In particular, the control unit 130 may include a detection unit 131, a calculation unit 132, and an image recognition unit 133.

The detecting unit 131 divides the test image into a plurality of blocks, acquires a plurality of regional feature descriptors for each divided block, and detects one local feature descriptor for each block of the test image using the object probability model . The detection unit 131 also detects a local feature descriptor for the reference image.

According to an embodiment of the present invention, the detection unit 131 may detect a local feature descriptor more than the reference image for a test image, and feature selection may be performed to select one local feature descriptor per block from the detected feature descriptor. Specifically, the directionality of the test image can be considered using the object probability model, and the detecting unit 131 'selects' a specific regional feature descriptor among the detected local feature descriptors according to the directionality. Therefore, the recognition performance can be enhanced even when the movement of the test image is changed.

The calculating unit 132 calculates the weight per block of the test image. For example, if the test image is a face image, there may be an occlusion region such as a face masked by an external object in the block of the face image. For such an area, the calculation unit 132 calculates the weight as low. Therefore, robust recognition results can be obtained despite the obstruction area.

The image recognition unit 133 sequentially compares the plurality of reference images with the test image, calculates similarity distance values between blocks matching each other using the selected regional feature descriptor and the weight, compares the calculated similarity distance values, And recognizes an image corresponding to the test image among a plurality of reference images. The image recognizing unit 133 can recognize the reference image having the smallest similarity distance value as the image matched with the test image. Alternatively, even if the similarity distance value is the smallest, the image recognition unit 133 can determine that the image is not recognized if the predetermined threshold value or more is exceeded.

The specific operations of the components included in the image recognition apparatus 100 are duplicated with the description of the above-described entity recognition method, and will not be repeated.

According to another embodiment of the present invention, the control unit 130 may include a determination unit 134 for determining whether to reject the test image.

Whether or not the image recognition unit 133 performs the recognition can be determined according to the determination result of the determination unit 134. [ For example, in the case where the image recognition apparatus 100 wants to recognize a face, if the input test image is not a face image but a landscape image, the similarity calculation with the reference image is meaningless. In this case, the efficiency of the image recognition apparatus 100 can be improved by rejecting the test image in advance.

According to an embodiment of the present invention, the determination unit 134 rejects the test image if the calculated value is less than the first threshold by applying the detected local feature descriptor to the object probability model. According to another embodiment, when the similarity distance value between the selected image and the test image is greater than the second threshold value, the determination unit 134 selects the image having the smallest similarity distance value from the test image among the plurality of reference images, . The various embodiments of the determination unit 134 described above may be applied independently to the image recognition apparatus 100, or may be simultaneously applied to exhibit synergy effects. As described above with respect to the image rejection method of the determination unit 134, detailed description is not provided.

FIG. 6 illustrates an example of images used in the image recognition method or image recognition apparatus according to an embodiment of the present invention.

6 (a), three test images of the user A 610, the user B 620, and the user C 630 are input.

For example, the test images of the user A 610 may include images having arbitrary movement in eight directions. Specifically, the left downwardly transformed image 611, the leftwardly deformed image 612, the leftwardly upwardly deformed image 613, the upwardly deformed image 614, 615, a right shifted image 616, a right downward shifted image 617, and a downwardly shifted image 618.

Even when a face detector boasts excellent performance, it is usual to always detect a face image to which such a moving deformation is applied. According to the above-described various embodiments, it is possible to consider the directionality due to the motion distortion of the test image, thereby providing a better recognition result than when the directionality is not considered. Referring to FIG. 7, the recognition accuracy shown by the graph 710 in the case where a weighted distance is taken into consideration and a step in which a local feature descriptor is selected is added Is higher than that of the graph 720 in the case of not otherwise. In the table of FIG. 7, the x-axis represents the degree of motion distortion of the test image, and the y-axis represents the recognition accuracy.

In Fig. 6 (b), the images that reflect the deformation that can be obtained in the actual situation are included. Specifically, the test images of (b) include a motion transformation as in (a), and a partial occlusion is also included. Experimental results including the test images of (b) can be seen from FIG. 8 shows a comparison example 1 (810) using a distance calculation function such as L ₁ norm, a comparison example 2 (820) in which SIFT and weight are considered, and an experimental example 830 in which SIFT, weight and feature selection are considered ) Were compared. As expected, it can be seen that the accuracy is higher in the experimental example (830).

The test images of FIG. 6 (c) include a face image 640 and images 650 that are not face images. Using the test images, the image rejection is performed through the determination unit 134 according to an embodiment of the present invention. As a result, it can be seen that the images 650 that are not the face images are effectively rejected, and consequently, the reliability of the face recognition is improved. The results are shown in Table 1 below.

[Table 1]

In this case, No rejection corresponds to a case in which no image rejection is introduced, which corresponds to a comparative example. In the case of the single rejection, when the above-described first image rejection is introduced, the dual rejection simultaneously introduces the first image rejection and the second image rejection . As shown in the table, it can be seen that the reliability of the image recognition in the case of the embodiment in which image rejection is introduced is doubled.

Meanwhile, the image recognition method according to various embodiments described above can be software-coded and stored in a non-transitory readable medium. Such non-transiently readable media can be used in various devices.

A non-transitory readable medium is a medium that stores data for a short period of time, such as a register, cache, memory, etc., but semi-permanently stores data and is readable by the apparatus. Specifically, it may be a CD, a DVD, a hard disk, a Blu-ray disk, a USB, a memory card, a ROM, or the like.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is to be understood that the invention is not limited to the disclosed exemplary embodiments, but, on the contrary, It will be understood by those skilled in the art that various changes and modifications may be made without departing from the spirit and scope of the present invention.

100: Image recognition device 110:
120: input unit 130:

Claims

In the image recognition method,
Dividing the test image into a plurality of blocks and acquiring a plurality of regional feature descriptors for each of the divided blocks when a test image is input;
Selecting one local feature descriptor for each block of the test image using an estimated object probability model based on a plurality of reference images;
Calculating a weight per block of the test image using the object probability model;
Calculating similarity distance values between blocks matching each other using the selected local feature descriptor and the weight value while sequentially comparing the plurality of reference images and the test image; And
Comparing the calculated similarity distance values and recognizing an image corresponding to the test image among the plurality of reference images.

The method according to claim 1,
Dividing the plurality of reference images into a plurality of blocks and obtaining one local feature descriptor for each block before a test image is input;
And estimating and storing the object probability model using the obtained local feature descriptors of the plurality of reference images when acquiring the local feature descriptor.

The method according to claim 1,
Wherein the obtained plurality of regional feature descriptors have different directional properties,
Wherein the selecting comprises:
Wherein one local feature descriptor corresponding to the directionality of the test image is selected for each block of the test image among the plurality of obtained local feature descriptors.

The method of claim 3,
Wherein the selecting comprises:

And selects the one local feature descriptor according to a formula such as < RTI ID = 0.0 >
Here, m is the index of each block,

(M = 1, ..., M) of m blocks of the test image,
And I ^tst is a test image.

The method according to claim 1,
When the local feature descriptor is selected and the weight is calculated, it is determined whether to reject the recognition of the test image using the one local feature descriptor selected for each block and the calculated weight per block Further comprising:
Wherein the step of calculating the similarity distance values and the step of recognizing the image are selectively performed according to the determination result of the determination step.

6. The method of claim 5,
Wherein,
Wherein the recognition of the test image is rejected if the value calculated by applying the selected local feature descriptor to the object probability model is less than a first threshold value.

6. The method of claim 5,
Wherein,
Selecting an image having the smallest similarity distance value from the test image among the plurality of reference images and rejecting recognition of the test image if the similarity distance value between the selected image and the test image is greater than a second threshold value Characterized in that:

3. The method of claim 2,
Wherein the calculating of the similarity distance values comprises:

And calculates the similarity distance values according to the following equation,
here

Is a local feature descriptor of m blocks of the ith reference image,
d () is the similarity distance calculation function,

A weight for the m block of the test image,
And I ⁱ is an i-th reference image.

An image recognition apparatus comprising:
A storage unit for storing an object probability model estimated using a regional feature descriptor for a plurality of blocks constituting each of a plurality of reference images;
An input unit for receiving a test image;
A detector for dividing the test image into a plurality of blocks, acquiring a plurality of regional feature descriptors for each divided block, and detecting one local feature descriptor for each block of the test image using the object probability model;
A calculating unit for calculating a weight per block of the test image using the object probability model; And
Comparing the plurality of reference images and the test image sequentially, calculating similarity distance values between blocks matching each other using the detected local feature descriptor and the weight, comparing the calculated similarity distance values, And an image recognition unit for recognizing an image corresponding to the test image among a plurality of reference images.

10. The method of claim 9,
Wherein the obtained plurality of regional feature descriptors have different directional properties,
Wherein:
For each block of the test image, one local feature descriptor corresponding to the directionality of the test image among the plurality of obtained local feature descriptors.

11. The method of claim 10,
Wherein:

And the local feature descriptor,
Here, m is the index of each block,

The k-direction local feature descriptor (k = 0, ..., 8) of m blocks of the test image,
p _m is the object probability model,

(M = 1, ..., M) of m blocks of the test image,
And I ^tst is a test image.

10. The method of claim 9,
Further comprising a determination unit configured to determine whether to reject the recognition of the test image,
Wherein the image recognizing unit selectively operates according to the determination result of the determining unit.

13. The method of claim 12,
Wherein,
And rejects recognition of the test image if a value calculated by applying the detected local feature descriptor to the object probability model is less than a first threshold value.

13. The method of claim 12,
Wherein,
Selecting an image having the smallest similarity distance value from the test image among the plurality of reference images,
And rejects recognition of the test image if a similarity distance value between the selected image and the test image is greater than a second threshold value.

10. The method of claim 9,
Wherein the image recognizing unit comprises: