KR20010087487A - Face recognizing method and apparatus using dimensional space transformation of Gabor filtering response - Google Patents

Abstract

PURPOSE: A method for recognizing a face using a dimensional space conversion of a gabor filtering reaction value and an apparatus therefor is provided to express the details of a face by using a gabor filtering reaction value as an input value for a dimensional space conversion. CONSTITUTION: A face detector(210) detects the position of a specific appearance of a face within an input image. A face area extractor(212) extracts a face area so as to maintain a constant rate with the detected position based on the position of the specific appearance of the face detected by the face detector(210). A size convertor(he214) converts at least one size into another size, so that the face area extracted by the face area extractor(212) and a gabor filter to be used have the constant rate to each other. An image preprocessor(216) reduces t effect of a lighting to the face area processed in the size convertor(214) by using a histogram smoothing process. A feature point extractor(218) extracts a predetermined number of the feature points in order to apply the gabor filter to them, in the face area pre-processed in the image preprocessor(216). A gabor filter(220) obtains a reaction value by applying a set of the gabor filters having various directions and frequencies to the respective feature point extracted by the feature point extractor(218). A matrix generating unit(230) expresses the reaction values obtained from the gabor filter(220) as one-dimensional array. A dimensional space converting unit(240) makes a new dimensional space corresponding to a matrix generated from the matrix generating unit(230) by using a predetermined dimensional space converting method. A database generating unit(250) maps each training image onto the new dimensional space and stores a coordinate value for each training image mapped onto the new dimensional space. A face similarity determining unit(260) transfers the one-dimensional array expressed by the matrix generating unit(230) to the new dimensional space already made by the dimensional space converting unit(240).

Description

Face recognizing method and apparatus using dimensional space transformation of Gabor filtering response

The present invention relates to face recognition, and more particularly, to a face recognition method and apparatus using dimensional transformation of Gabor filtering response values.

As the modern society rapidly develops into an information society, a lot of information is pouring out. Systems that handle this information allow only specific users to share the information when needed, and block access to other users. Many systems for information security have been invented. Information can be not only some data, but also a spatial concept of a particular place. In addition to existing keys and IC cards, various security devices have been invented. However, as there are many more places to access, there are also a number of tools for revoking this security.

As one method for solving these shortcomings, research is being actively conducted on security methods using the physical characteristics of a person. The advantage of this method is that it can't be easily replicated and you don't have to carry any tools. As a security method using physical features, there are technologies using fingerprints, irises, faces, and voices. In particular, the feature of the method using the face is that the user does not require any specific behavior when the user wants to access the system.

Therefore, face recognition technology has been studied a lot, and representative methods include face recognition method using principal component analysis (PCA) and Gabor filter.

The face recognition method using the PCA recognizes the face by the composition of the overall face rather than looking at the detailed face of the face, and has the following disadvantages. For the frontal face image to be used as input, the face must be exactly aligned, and its size must be normalized to some extent. On the other hand, there is a disadvantage that shows a sensitive response to changes in lighting and brightness. The face recognition method using PCA uses real face images as inputs and constructs a new space based on these images, and thus is sensitive to changes in illumination and brightness due to limitations of the input images.

The face recognition method using Gabor filters applies a set of 2-D Gabor filters having different frequencies and directions to a predetermined region of the face based on the feature points of the face, and identifies the face from the response. The disadvantages of revealing weaknesses in grasping the composition of the entire area of the face at once and requiring a relatively large amount of storage space for storing personal data.

The technical problem to be achieved by the present invention is to solve the above-mentioned problem, and to move the Gabor filtering response value obtained from a certain region of the face based on the feature points of the face to the new dimension space through the process of dimension space conversion. The present invention provides a face recognition method using a dimensional space transformation of Gabor filtering response values to increase face recognition performance and reduce the amount of database by performing face recognition after the face recognition.

Another object of the present invention is to provide a face recognition apparatus using dimensional space transformation of Gabor filtering response values for performing the face recognition method.

1 is a flowchart illustrating a face recognition method according to the present invention.

2 is a block diagram according to a preferred embodiment of the face recognition apparatus according to the present invention.

FIG. 3 is a diagram illustrating an example of extracting a face region performed by the face region extractor illustrated in FIG. 2.

4 is a diagram illustrating an example of extracting a facial feature point performed by the facial feature point extractor illustrated in FIG. 2.

FIG. 5 is a diagram illustrating an example of a face recognition process performed by the apparatus shown in FIG. 2.

In order to achieve the above object, the face recognition method according to the present invention,

(a) For the training images to be referred to for face recognition, map Gabor filter response values obtained from the feature points of the face to the new dimensional space created through the dimensional space transformation process, and map each training image in the new dimensional space. Generating a face model database by storing the coordinate values for the face and (b) moving the Gabor filter response values obtained from the feature points of the face to a new dimensional space for the test image to perform face recognition. Identifying a similar face for the test image by referring to the model database.

In order to achieve the above another object, the face recognition apparatus according to the present invention,

A Gabor filter unit that obtains response values by applying a set of Gabor filters having various directionalities and frequencies to each feature point of the face extracted from the input image, and expresses the response values in a one-dimensional array. A matrix generator for constructing each one-dimensional array obtained for each training image into a matrix when inputted as a dimensional space; a dimensional space converter for creating a new dimensional space corresponding to a matrix by using a predetermined dimensional space transform method; When the test image is input to the face model database generation unit and the input image that maps the training image to the new dimensional space and stores the coordinate values for each training image in the new dimensional space, the Gabor filter unit and the matrix generation unit One-dimensional array represented through the new dimension space And a face similarity determining unit for identifying a similar face by referring to the face model database.

Hereinafter, a face recognition apparatus and a method using a dimensional space transformation of Gabor filtering response values according to the present invention will be described with reference to the accompanying drawings.

1 is a flowchart illustrating a face recognition method according to the present invention.

The present invention utilizes only the features of the face recognition method using Gabor filter and the features of the face recognition method using PCA. Using these two methods, the face recognition method according to the present invention greatly generates a face model database, and recognizes a face of an input image to be actually tested by referring to the face model database.

To generate a face model database, training images to be referred to for face recognition are input. For each training image, Gabor filtering response values obtained from facial feature points are mapped to a new dimensional space created through the dimensional space transformation process, and the coordinate values of each training image in the new dimensional space are stored.

In detail, referring to FIG. 1, response values are obtained by applying a set of Gabor filters having various directionalities and frequencies to each of the feature points of the face extracted from each of the sequentially input training images (step 100). . Here, the feature points use parts sensitive to Gabor filtering response in the face region through experiments. Experimentally, response values are obtained by applying a set of Gabor filters to a region based on a feature point in a face region.

A process for extracting the feature points necessary in step 100 is briefly as follows. First, the position of a specific feature of a face is found in a training image including a face image. For example, the position of the binocular of the face, the overall shape of the face, and the like. Next, the face region is extracted from the training image using the position of the face specific feature. After changing the size of the face region or Gabor filter so that the extracted face region and Gabor filter have a certain ratio, the histogram smoothing process is performed on the face region to reduce the influence of illumination. Next, feature points to be applied to the Gabor filter in the face region are extracted.

After operation 100, the response values obtained by applying the Gabor filters are represented as a one-dimensional array, and the one-dimensional arrays obtained for each training image are configured as one matrix (operation 102). That is, the response values obtained from one training image are changed into a one-dimensional array, and the same method is applied to the other training images to change the response values into a one-dimensional array. Since each feature has a plurality of response values for one training image, for example, if there are 20 feature points and a set of 40 Gabor filters is used, 800 response values will be obtained for one training image. This is expressed as a one-dimensional array, and each one-dimensional array represented in this way constitutes a matrix constituting one column or row. At this time, the matrices are composed of a covariance matrix or a correlation matrix again to reduce the size.

After step 102, a new dimensional space corresponding to the matrix constructed in step 104 is created using a predetermined dimensional space transformation method (step 104). As the method of transforming the dimensional space, that is, moving to another dimension space, preferably, a principal component analysis (PCA) method, an independent component analysis method, and a nonlinear PCA method are used. Can be.

When a new dimensional space is created, each training image is mapped to a new dimensional space (step 106). Next, a coordinate model in the new dimensional space is stored to generate a face model database (step 108). At this time, the class is composed of face images of the same person among the training images moved to the new dimensional space. The new dimensional space reconstructs the matrix to distribute images having similar facial shapes or specific features of similar faces.

After the face model database is generated through the steps 100 to 108, the face image may be input to perform face recognition. Even for the test image, Gabor filter response values obtained from facial feature points are moved to a new dimension space. Thereafter, a similar face is identified for the test image by referring to the face model database.

In detail, referring to FIG. 1, response values are obtained by applying a set of Gabor filters having various directionalities and frequencies to each of feature points of a face extracted from an input real test image (step 110). Here, the feature points use parts sensitive to Gabor filtering response in the face region through experiments. The procedure for extracting the feature points is also performed as it is performed for the training image.

After operation 110, the response values obtained by applying the Gabor filters are expressed as a one-dimensional array (operation 112). The one-dimensional array is moved to a new dimensional space created using a predetermined dimensional space transformation method in step 104 (step 114). A similar face for the text image is identified by referring to the face model database (step 116). That is, it identifies which class the test image is included in in the new dimension space.

2 is a block diagram according to a preferred embodiment of the face recognition apparatus according to the present invention.

The face recognition apparatus according to the present invention includes a face detector 210, a face region extractor 212, a size converter 214, an image preprocessor 216, and a face feature point extractor 218 as preprocessors for Gabor filtering. And a Gabor filter 220, a matrix generator 230, a dimensional space converter 240, a face model database generator 250, and a face similarity determiner 260.

The face detector 210 detects a position of a specific feature of the face in the input image. For example, it may be set to detect the position of both eyes as the position of a specific appearance of the face. If a face image is included in the input image, the position of two eyes having a distance within a predetermined range may be found.

The face region extractor 212 extracts an area of the face to maintain a constant ratio with the position based on the position of a specific feature of the face detected by the face detector 210. For example, if the face detector 210 detects the positions of the two eyes, the distance of the two eyes may be obtained. As a preprocessing step for face recognition, a method of cutting out a face region from an input image based on a distance between two eyes is to minimize an influence on a background of an image or a change in a human hair style.

3 illustrates an example of extracting a face region performed by the face region extractor 212 illustrated in FIG. 2, 300 represents an input image, and 302 represents an extracted face region.

For example, when the distance between two eyes is 2D, the distance outside the left eye and the distance outside the right eye are respectively D, and the area distance above the eyes is 1.5 * D relative to the eyes, and the eyes is relative to the eyes. A 3 * D area distance under the eye is trimmed to remove the face area suitable for face recognition. Referring to FIG. 3, it can be seen that when the face area is cut at such a ratio, the background is minimized while having a simple numerical value, and the face features are included in the entire face such as eyebrows, eyes, nose, and lips.

Referring back to FIG. 2, the size converter 214 converts at least one size such that the face region extracted by the face region extractor 212 and the Gabor filter to be applied later have a predetermined ratio with each other. In one method, the size of the face area is changed according to the size of the 2-D Gabor filter to be applied later. That is, the size of the image is changed to maintain a constant ratio in consideration of specific features of the face area, for example, the distance between the two eyes and the size of the 2-D Gabor filter. Alternatively, the ratio of Gabor filters may be varied to maintain a constant ratio. Here, the size change of the Gabor filter means a change in resolution.

The image preprocessor 216 reduces the influence of illumination by performing histogram smoothing on the face region passed through the size converter 214. The input images may have a relatively high or low luminance due to the illumination, and a high portion and a low portion may be generated even in one input image. In order to reduce the influence of illumination, the histogram is obtained by analyzing the luminance distribution of each pixel in the face region, and the histogram is smoothed around the high frequency luminance.

The facial feature point extractor 218 extracts a predetermined number of feature points to which the Gabor filter is to be applied from the face region preprocessed by the image preprocessor 216. 4 is a diagram illustrating an example of extracting a facial feature point performed by the facial feature point extractor illustrated in FIG. 2. Referring to FIG. 4, in this experiment, for example, 20 feature points are displayed in a cross shape. The criterion indicated above is the result of experiments on the sensitive parts of Gabor filter in the face area.

The Gabor filter 220 obtains response values by applying a set of Gabor filters having various directions and frequencies to each of the facial feature points extracted by the facial feature extractor 218. In general, the response from the Gabor filter ( ) Is obtained from Equation 1 below.

In Equation 1, Denotes frequency and directivity, respectively, in this experiment, the range of ν is 0 to 4, and the range of μ is 0 to 7. Therefore, a total of 40 Gabor filters were used.

FIG. 5 is a diagram illustrating an example of a face recognition process performed by the apparatus shown in FIG. 2. Referring to FIG. 5, 500 denotes an input image, 502 denotes a face region extraction result from the face detector 210 and the face region extractor 210, and 504 denotes a size converter 216, an image preprocessor 216, and the like. Preprocessing and feature point extraction results from the facial feature point extractor 218 are shown, respectively.

506 is a characteristic response extraction result of applying 16 Gabor filters having different directions and frequencies to one feature point of the face region. In this experiment, Gabor filters were applied to a specific area of a certain size around the feature points. For each Gabor filter as shown in 506, response values for a predetermined area of a predetermined size are obtained. As a result, one response value is extracted and used.

Referring back to FIG. 2, the matrix generator 230 expresses the response values obtained by the Gabor filter 220 in a one-dimensional array. In order to generate a face model database, if training images are sequentially input as input images, one-dimensional arrays obtained for each training image are configured as one matrix. Gabor filtering response values obtained from one training image are expressed in a one-dimensional array, and the same method is applied to the remaining training images to form a matrix in which each row or column is formed. This matrix is converted back into a covariance matrix or a correlation matrix.

The dimensional space converter 240 creates a new dimensional space corresponding to the matrix constructed by the matrix generator 230 using a predetermined dimensional space transformation method. Here, the dimensional space transformation method may be preferably used as a principal component analysis (PCA) method, independent component analysis (ICA) method and nonlinear principal component analysis method. Typically, principal component analysis (PCA) methods feature a set of eigen vectors in a multi-dimensional image space.

The database generator 250 maps each training image to a new dimensional space, and then stores coordinate values of each training image in the new dimensional space. At this time, the class is composed of the face images of the same person among the training images moved to the new dimensional space.

Referring to FIG. 5, 508 represents a matrix constructed in the matrix generator 230. One column in the matrix corresponds to one training image, and k in one column becomes 800 if, for example, 40 Gabor filters were applied for 20 feature points. However, k is reduced to 400 using a communicative matrix or correlation matrix. As a result of performing the feature coefficient dimension transformation on this matrix, a new dimension space is created, one column of the matrix being a coordinate value in the new dimension space. The face database is mapped to these coordinate values as shown in 510.

When training images are sequentially input to the input image, a face model database is generated from the components described so far. Meanwhile, when the test image is actually input as an image to be newly recognized, that is, an input image, the block 200 indicated by a dotted line in FIG. 2 is performed in the same manner as that performed for the training image.

Next, the matrix generator 230 expresses the response values obtained from the Gabor filter 220 in a one-dimensional array with respect to the test image. The face similarity determination unit 260 moves the one-dimensional array expressed through the matrix generator 230 to the new dimensional space already created by the dimensional space converter 240. The database generator 250 identifies which classes are included in the new dimension space and whether there are similar faces. Finally, the face recognition result is output.

The face recognition method and apparatus according to the present invention described so far can be applied to an application requiring a face recognition technique. For example, when applied to a security device, access can be managed by using the physical characteristics of the individual instead of a key or IC card where restricted access is required.

In addition, when applied to the face database search apparatus, it is possible to find a specific person among many people, such as searching a criminal list. In addition, when applied to the attendant counting system, it is possible to check whether people in a closed space attend. It can be applied to many other applications.

As described above, the face recognition method and apparatus according to the present invention express the detailed faces of the face by using Gabor filtering response as an input for dimensional space transformation, and change in rotation and illumination due to the characteristics of Gabor filter. This feature has the advantage of significantly reducing the size of the face database in other spaces by increasing the face recognition performance by applying strong characteristics to the space and applying the dimensional space transformation method.

Claims (17)

(a) For the training images to be referred to for face recognition, map Gabor filter response values obtained from the feature points of the face to the new dimensional space created through the dimensional space transformation process, and map each training image in the new dimensional space. Generating a face model database by storing coordinate values for the face; And (b) For the test image to perform face recognition, move Gabor filter response values obtained from the feature points of the face to the new dimensional space, and then refer to the face model database for a similar face for the test image. Face recognition method comprising the step of identifying. According to claim 1, wherein the step (a), (a1) obtaining response values by applying a set of Gabor filters having various directions and frequencies to each of the feature points of the face extracted from each of the sequentially input training images; (a2) expressing the response values in a one-dimensional array and constructing one-dimensional arrays obtained for each training image into a matrix; (a3) creating a new dimensional space corresponding to the matrix using a predetermined dimensional space transformation method; And (a4) mapping each training image to a new dimensional space and storing coordinate values for each training image in the new dimensional space to generate a face model database. The method of claim 1 or 2, wherein step (b) comprises: (b1) obtaining response values by applying a set of Gabor filters having various directions and frequencies to each of the feature points of the face extracted from the input test image; (b2) expressing the response values in a one-dimensional array; (b3) moving the one-dimensional array to the new dimensional space; And (b4) identifying a similar face with respect to the test image by referring to the face model database. The method of claim 2, wherein in the step (a2), the matrix, And the one-dimensional array is a covariance matrix for a matrix constituting one column or row. The method of claim 2, wherein in the step (a2), the matrix, And the one-dimensional array is a correlation matrix for a matrix constituting one column or row. The method of claim 2, wherein in the step (a3), the predetermined dimensional space transformation method, Face recognition method characterized in that the principal component analysis (PCA) method. The method of claim 2, wherein the predetermined dimensional space transformation method in the step (a3) is an independent component analysis (ICA) method. The method of claim 2, wherein in the step (a3), the predetermined dimensional space transformation method, Face recognition method characterized in that the non-linear principal component analysis method. A Gabor filter unit for obtaining response values by applying a set of Gabor filters having various directions and frequencies to each of the feature points of the face extracted from the input image; A matrix generator configured to express the response values in a one-dimensional array and configure one-dimensional arrays obtained for each training image as one matrix when the training images are sequentially input to the input image; A dimension space transform unit which creates a new dimension space corresponding to the matrix by using a predetermined dimension space transform method; A face model database generator for mapping each training image to the new dimensional space and storing coordinate values for each training image in the new dimensional space; and When a test image is input to the input image, the one-dimensional array represented by the Gabor filter unit and the matrix generator is moved to the new dimensional space, and the similar face is identified by referring to the face model database. Face recognition apparatus comprising a face similarity determination unit. The apparatus of claim 9, wherein the face recognition device comprises: A face sensor for detecting the position of a specific feature of the face in the input image; A face region extractor extracting a face region based on the position; A size converter for converting at least one size so that the extracted face region and the Gabor filter have a predetermined ratio with each other; And And a face feature point extractor for extracting a predetermined number of feature points to be applied to the Gabor filter in the extracted face region. The apparatus of claim 10, wherein the face recognition device comprises: And an image preprocessor configured to reduce the influence of illumination on the extracted face region to be input to the facial feature point extractor through a histogram smoothing process. The method of claim 9, wherein the face model database generation unit, After each training image is moved to a new dimensional space, the face recognition device characterized in that to find a space in which the face image of the same person among the training images are distributed and classify the class in the new dimensional space. The method of claim 9, wherein the matrix, And the one-dimensional array is a covariance matrix for a matrix constituting one column or row. The method of claim 9, wherein the matrix, And the one-dimensional array is a correlation matrix for a matrix constituting one column or row. The method of claim 9, wherein the predetermined dimensional space transformation method, A facial recognition apparatus, characterized in that the principal component analysis (PCA) method. The method of claim 9, wherein the predetermined dimensional space transformation method, Face recognition apparatus, characterized in that the independent component analysis (ICA) method. The method of claim 9, wherein the predetermined dimensional space transformation method, A face recognition apparatus, characterized in that the nonlinear principal component analysis method.