KR20100073749A - Apparatus and method for extracting feature point based on sift, and face recognition system using thereof - Google Patents

Abstract

PURPOSE: A feature point extraction apparatus based on S SIFT(Scale Invariant Feature Transform) capable of improving visual recognition rate by increasing the number of feature points is provided to increase the number of extracted keypoints by calculating optimized contrast threshold. CONSTITUTION: A DOG image generator unit(710) generates a DOG image with the Gaussian pyramid of the input image. A candidate pixel extractor unit(720) extracts the sample point corresponding to maximum or the minimum of each DOG image. A contrast threshold calculator(730) calculates contrast threshold of candidate pixels used in feature point extraction responding to the input image.

Description

Apparatus and method for extracting feature point based on SIFT, and face recognition system using

The present invention relates to an image recognition technology based on SIFT, and in particular, a feature point extraction apparatus and method that is robust to a fluid environment by calculating an optimized density threshold of candidate pixels used for feature point extraction according to an input image, and a face recognition system using the same. And to a method.

Recently, a technique of extracting a feature point from an image acquired by using a camera has been widely used in computer vision fields such as personal authentication through face recognition, 3D reconstruction, and self-tracking.

The feature point of the image should be invariant to the change in size and rotation of the image. In addition, it must be partially invariant to lighting and camera pose changes.

Conventional methods using Principal Component Analysis (PCA) or Independent Component Analysis (ICA) have high recognition rates for general images with many edges, but they are not suitable for face images and endoscopy images that are difficult to extract feature points. There was a problem.

In particular, the existing methods have a satisfactory level of face recognition in a limited situation, but have a problem that is difficult to apply in a fluid situation. For example, there is a problem that the recognition rate is remarkably decreased for images having different sizes, images including not only faces but also backgrounds.

Accordingly, the first technical problem to be achieved by the present invention is to increase the number of extracted feature points and improve the recognition rate by calculating an optimized density threshold of candidate pixels used for feature point extraction based on the input image based on SIFT. It is to provide a feature point extraction apparatus that can be.

A second technical problem to be achieved by the present invention is to provide a face recognition system using the feature point extraction apparatus.

The third technical problem to be achieved by the present invention is to calculate the optimal density threshold value of candidate pixels used for feature point extraction according to the input image based on SIFT, thereby increasing the number of extracted feature points and improving the recognition rate. It is to provide a feature point extraction method.

The fourth technical problem to be achieved in the present invention is to provide a face recognition method using the feature point extraction method.

In order to solve the first technical problem as described above, the present invention, in the apparatus for extracting the feature point of the input image based on the SIFT, the concentration threshold value of the candidate pixel (candidate pixel) used to extract the feature point corresponding to the input image a concentration threshold calculator for calculating a contrast threshold; And a feature descriptor descriptor generating unit for generating a keypoint descriptor by using the candidate pixels having the density threshold or higher among all the candidate pixels.

In one embodiment, the concentration threshold calculator calculates a density value having the maximum number of pixels as a density threshold value through a histogram analysis of a difference of Gaussian (DOG) image generated from the input image.

In one embodiment, the feature point descriptor generator generates the feature point descriptor using feature point information including position, size, and orientation information of a neighboring pixel based on the candidate pixel above the density threshold.

In one exemplary embodiment, the apparatus for extracting a feature point may include: a DOG image generator configured to generate a DOG image using a Gaussian pyramid of the input image; And a candidate pixel extracting unit extracting the full candidate pixel by searching for the maximum or minimum point of the DOG image.

In order to solve the second technical problem as described above, the present invention, based on the SIFT, calculates the contrast threshold of the candidate pixel (candidate pixel) used for feature point extraction corresponding to the input face image, A feature point extracting unit extracting feature points using candidate pixels equal to or greater than the density threshold value; And a face recognition unit configured to recognize the input face image by matching the feature points of the input face image with the feature points of the face image stored in the database.

In one exemplary embodiment, the feature point extractor is configured to calculate a density value having a maximum number of pixels as the concentration threshold value through a histogram analysis of a difference of Gaussian (DOG) image generated from the input face image. A threshold calculator; And a feature descriptor generator for generating a keypoint descriptor by using candidate pixels having the density threshold or higher among all candidate pixels.

In one embodiment, the face recognition unit matches feature points of the input face image and the stored face image through clustering by a Hough Transform.

In one embodiment, the face recognition unit recognizes the input face image in consideration of the number and distribution information of the matching feature points.

In order to solve the third technical problem as described above, the present invention provides a method for extracting feature points of an input image based on a SIFT in a computer system, wherein the candidate pixels used for feature point extraction corresponding to the input image are selected. A concentration threshold calculating step of calculating a contrast threshold; And a feature point descriptor generation step of generating a keypoint descriptor by using the candidate pixels having the density threshold or higher among all the candidate pixels.

In order to solve the fourth technical problem as described above, the present invention provides a method for recognizing a face image based on a SIFT in a computer system, wherein the candidate pixel is used for extracting feature points based on the SIFT and corresponding to the input face image a feature point extraction step of calculating a contrast threshold of a candidate pixel and extracting feature points using candidate pixels equal to or greater than the concentration threshold; And a face recognition step of matching the feature point of the input face image with the feature point of the face image stored in the database to recognize the input face image.

SIFT-based feature point extraction apparatus and method and face recognition system and method using same according to the present invention, by calculating the optimal density threshold value of the candidate pixel used for feature point extraction according to the input image the number of feature points extracted than the conventional method It increases the number by more than 30%, thereby improving the recognition rate.

Prior to the description of the specific contents of the present invention, for the sake of understanding, an outline of the technical problem to be solved by the present invention is first presented.

First, the present invention is based on the Scale Invariant Feature Transform (SIFT) algorithm to create a face recognition system that is robust to a fluid environment.

SIFT, a representative algorithm for extracting image feature points, is described in DG Lowe, "Distinctive image features from scale-invariant keypoints", Int. J. Comput. Vision 60 (2), 91-110, 2004.]. SIFT is invariant to image rotation, scaling, shifting, partial illumination changes, and projective transforms. SIFT extracts attributes such as the position, scale, and direction of a feature in consideration of local image characteristics. That is, in the first step, the sample point (or candidate pixel) is selected by searching the maximum and minimum in the scale space generated by the difference-of-Gaussian (DOG) function ( Scale - space extrema detection ) . In the second step, keypoints are selected based on stability values ( Keypoint localization ) . The third step, it allocates at least one direction with respect to each key point (Orientation assignment ) . Finally, create a keypoint descriptor using local image gradients ( Keypoint descriptor ) .

In particular, SIFT is used to find a fundamental matrix (F), which is the first step in finding three-dimensional information in face recognition. That is, feature points invariant to the change in size are extracted by SIFT, and the base matrix F is obtained using the feature points. In the process of obtaining the basic matrix F from the feature points, clustering by keypoint matching and Hough Transform to remove outliers and solve a one-to-one correspondence problem. Go through the (clustering) step.

Through this process, the number of matching points and distribution information between two images are identified to recognize a face.

However, since a simple SIFT algorithm applies a reference value for extracting a feature point to any image, there is a limit in that the number of feature points that can be extracted is small for a face image, an endoscope image, etc. having less features.

FIG. 1 illustrates feature points extracted by applying SIFT to two general images.

2 illustrates a feature point extracted by applying SIFT to two endoscope images.

1 and 2 illustrate a case where a contrast threshold of a candidate pixel used for feature point extraction is set to 0.02 and a curvature threshold is set to 10. FIG.

As shown in FIG. 1, the 485 and 414 feature points are extracted for the general image from the left image, respectively.

However, as shown in FIG. 2, 18 and 13 feature points are extracted from the left image, respectively, for the endoscope image, indicating that there is a limitation in application for the image having few features.

In contrast, the present invention increases the number of extracted feature points and improves the recognition rate by properly calculating a reference value of feature point extraction according to an input image in order to overcome the limitation of SIFT.

3 illustrates a feature point extracted by applying a SIFT set to a concentration threshold value of 0.01 and a curvature threshold value of 10 for two endoscope images.

As shown in FIG. 3, it can be seen that performance is improved by extracting 111 and 110 feature points from the left image, respectively.

FIG. 4 illustrates feature points extracted by applying SIFTs having a concentration threshold value of 0.01 and a curvature threshold value of 50 for two endoscope images.

As shown in FIG. 4, it can be seen that the feature points of the point shown as the boundary are slightly increased by extracting 126 and 123 feature points from the left image, respectively.

As can be seen from the results of FIGS. 3 and 4, the performance of the SIFT is significantly affected by the concentration threshold compared to the curvature threshold.

Therefore, the present invention is particularly intended to increase the number of extracted feature points by automatically calculating an optimized density threshold of candidate pixels used for feature point extraction according to an input image, and to improve face recognition rate.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings in order to clarify the solutions of the technical problems of the present invention. However, in describing the present invention, when it is determined that the detailed description of the related known technology or configuration may make the gist of the present invention unclear, the detailed description thereof will be omitted. In addition, terms to be described later are terms defined in consideration of functions in the present invention, which may vary according to intention or custom of a user, an operator, or the like. Therefore, the definition should be made based on the contents throughout the specification.

5 is a block diagram illustrating an example of a face recognition system based on SIFT according to the present invention.

6 is a flowchart illustrating an example of a face recognition method based on SIFT according to the present invention.

5 and 6, the face recognition system 500 according to the present invention includes a feature point extractor 502 and a face recognizer 504, and include a camera unit 510 and an image signal processor 520. And a face area detector 530.

First, the camera unit 510 converts an optical signal of a subject into an electrical signal using an image sensor, for example, a CCD sensor or a CMOS sensor (S610).

The video signal processor 520 converts the electrical analog video signal input from the camera unit 510 into a digital data video signal (S620).

The face region detector 530 searches for and extracts a face region from the image signal input from the image signal processor 520 (S630). In order to implement the face region detector 530, a face extraction technique using an outline, a skin color, a texture, a template, or the like may be used. For example, the face area detector 530 may find a face area using a face color in an input image. In this case, the face region detection unit 530 may include a transformation module for emphasizing face color, a transformation module for emphasizing black and white, a preprocessing module for enhancing an image of a specific portion to extract facial feature information, and a face region; And a background labeling module for separating the background area.

Next, the feature point extractor 502 calculates a contrast threshold of a candidate pixel used for feature point extraction based on an SIFT but corresponding to an input face image, and calculates the density threshold. The feature point is extracted using the candidate pixels as described above (S640).

The feature point extractor 502 may be implemented by a feature point extraction apparatus and method according to the present invention.

7 is a block diagram of an SIFT-based feature point extraction apparatus according to the present invention.

8 is a flowchart illustrating a feature extraction method based on the SIFT according to the present invention.

7 and 8, the apparatus for extracting feature points according to the present invention includes a DOG image generator 710, a candidate pixel extractor 720, a density threshold calculator 730, and a feature descriptor generator 740. Include.

The DOG image generator 710 generates a DOG image by using a Gaussian pyramid of the input image (S810).

9 illustrates a process of generating a DOG image using a Gaussian pyramid of an input image.

As illustrated in FIG. 9, the DOG image generator 710 passes the input image through a Gaussian filter having dispersion values of various scales, and as a result, the images are sequentially subtracted from each other according to adjacent scales. The resulting image is called a Gaussian Difference (DOG) image. The DOG image is subsampled 2: 1 and passed through a Gaussian filter with various scales. Through this process, DOG images of pyramid structure can be obtained.

Next, the candidate pixel extracting unit 720 extracts all candidate pixels to be sample points, that is, feature points, corresponding to the maximum or minimum of each DOG image (S820).

10 shows pixels of a DOG image compared to extract candidate pixels.

As shown in FIG. 10, the maximum or minimum position is extracted as the candidate pixel when the maximum or minimum is 8 pixels and 9 pixels of the up and down scale DOG image, both of which are maximum or minimum compared to 26 pixels.

Next, the density threshold calculator 730 calculates a contrast threshold of a candidate pixel used for feature point extraction in response to the input image (S830).

In the conventional SIFT, candidate pixels on the edge line having a small or strong DOG value are removed to remove an element due to noise. That is, candidate pixels smaller than the density threshold value are removed as in Equation 1.

,

,

,

,

,

,

.

.

는 localization을 통해 정확성이 향상시킨 위치 정보, 그리고

는 localization된 픽셀에 대한 DOG를 의미한다.In Equation 1, D is a variable representing a DOG, X is information on the position of the selected pixel,

Location information has improved accuracy through localization, and

Means DOG for localized pixels.

As shown in Equation 1, the smaller the concentration threshold value, the more feature points can be extracted. However, when the density threshold is randomly reduced to increase the number of extracted feature points more than necessary, problems such as inefficiency of calculation and risk of mismatching occur.

값의 분포가 다르므로 입력 영상에 따라 적절한 농도 임계점을 산출할 필요가 있다.In addition, every input video

Since the distribution of the values is different, it is necessary to calculate an appropriate concentration threshold according to the input image.

However, the existing SIFT has a problem of extracting feature points using fixed reference values regardless of the input image.

In contrast, the concentration threshold calculator 730 of the present invention measures the density value having the maximum number of pixels through histogram analysis of a difference of Gaussian (DOG) image generated from the input image. Calculate as

The histogram analysis refers to a method of analyzing a number of pixels having a corresponding density level or a ratio of all pixels in an image for each level of density. The results of such histogram analysis can be generally displayed as a bar graph in which the concentration value is taken along the horizontal axis and the number of pixels along the vertical axis. The histogram analysis may be used to examine what density values the image data are composed of.

Meanwhile, the condition of the candidate pixel removed in relation to the curvature threshold r may be derived as in Equation 2.

As shown in Equation 2, in the case of the curvature threshold value, as the value of r increases, the number of feature points extracted from the portion shown as the boundary may be increased slightly.

Next, the feature point descriptor generator 740 generates a keypoint descriptor using candidate pixels equal to or greater than the density threshold value among all candidate pixels (S840).

The feature point descriptor generator 740 generates the feature descriptor using the feature point information including the position, the size, and the orientation information with respect to the neighboring pixel based on the candidate pixel above the density threshold. That is, a 16x16 block is set around the last selected candidate pixel, and the directionality of the pixels within the block is measured. The direction and the size of each pixel are measured to estimate the direction that is largest in the block. This direction is set to the direction of the feature point. In order to obtain the final SIFT feature vector, 16x16 blocks are set around each key point, and then divided into 4x4 block units to measure the direction of change of pixel value for each pixel in the block. The change direction is set to be proportional to the change slope to obtain a probability distribution with respect to the direction so that the larger the magnitude of the change, the greater the probability distribution of the corresponding direction.

FIG. 11 illustrates a process of extracting directional distributions for 8 by 8 blocks and quantizing them in eight directions.

As shown in FIG. 11, the directional probability distribution is stored in eight directions, which are represented in a 128-dimensional vector form for a 16 × 16 block. The vector representing the directional probability distribution is a structure of the SIFT feature vector, and is stored together with parameters such as image coordinates, directionality, and Gaussian scale of the feature point. The 128-dimensional SIFT feature vector thus obtained is represented by a codebook by performing vector quantization. The initially obtained SIFT feature vector is any real number with floating point, which requires a relatively large capacity and time when transmitting the SIFT feature vector or comparing distances. Therefore, in order to overcome this disadvantage, vector quantization is performed on 8-dimensional vectors of SIFT divided into 8-direction units, and 8-dimensional direction distributions expressed by real values are mapped to 1-dimensional codebooks. Through this vector quantization process, a 128-dimensional real coefficient SIFT feature vector can be converted into a 16-dimensional integer coefficient vector. When the SIFT feature vector quantized with the 16-dimensional vector is transmitted with a low data amount and the difference between the vectors is calculated, the SIFT feature vector is reduced and compared with the 128-dimensional vector of the quantized codebook.

As described above, when the feature point extractor 502 extracts the feature point from the input face image, the face recognition unit 504 matches the feature point of the input face image with the feature point of the face image stored in the database 540. The input face image is recognized (S650).

That is, the face recognizing unit 504 matches feature points of the input face image and the stored face image through clustering by a Hough transform, and uses the matching feature points to match the feature points. An image having a large number of feature points to be recognized is recognized as a face image. In addition, the face recognizing unit 504 may extract the face image in consideration of the distribution information of the matching feature points such as the location of the matching feature points, the length and the slope of the line segments connecting the matching feature points, as well as the number of matching feature points. I can recognize it.

In particular, when the face recognition unit 504 has face image information having a matching point equal to or greater than a predetermined reference value among the face image information stored in the database 540 in personal authentication and system security, It may include an authentication module (not shown) that recognizes the access rights, otherwise denies the access rights.

In one embodiment, the SIFT-based feature extraction apparatus and face recognition system using the same is implemented in the form of a system on chip on a microprocessor to allow the microprocessor to extract and Face recognition can be performed. If the present invention is implemented in a microprocessor, the size of various systems can be reduced, the assembly process can be simplified, and manufacturing costs can be reduced.

On the other hand, the detailed process of the SIFT-based feature point extraction method and the face recognition method using the same according to the present invention can be described corresponding to the detailed operation of the configuration module of the feature point extraction apparatus and the face recognition system.

In addition, the SIFT-based feature point extraction method and the face recognition method using the same can be implemented as a program code that can be read by a computer in a computer-readable recording medium. When the present invention is executed through software, the constituent means of the present invention are code segments for performing necessary tasks. The program or code segments may be stored on a processor readable medium or transmitted by a computer data signal coupled with a carrier on a transmission medium or network.

The computer-readable recording medium includes all kinds of recording devices for storing data that can be read by a computer system. Examples of computer-readable recording media include ROM, RAM, CD-ROM, magnetic tape, floppy disks, optical data storage devices, and the like. The computer readable recording medium can also be distributed over network coupled computer systems so that the computer readable code is stored and executed in a distributed fashion.

Hereinafter, the significant effects of the present invention are verified.

12A and 12B illustrate feature points and matching points extracted by applying the present invention to a general image, respectively.

As shown in FIGS. 12A and 12B, the number of extracted feature points was 687 and 772, respectively, from the left side, and the calculated concentration thresholds were 0.0077 and 0.0056, respectively, from the left side. In addition, 135 matching points were obtained.

13A and 13B illustrate feature points and matching points extracted by applying the present invention to an endoscope image, respectively.

As shown in FIGS. 13A and 13B, the number of extracted feature points was 560 and 509, respectively, from the left side, and the calculated concentration thresholds were 0.0031 and 0.0032, respectively, from the left side. In addition, there were 145 matching points.

As described above, according to the present invention, it is understood that more feature points can be extracted regardless of the type of image, and thus more matching points can be found.

14A to 18B illustrate face recognition results using the conventional method and the present invention.

14A and 14B show feature points extracted by applying the conventional method and the present invention, respectively.

15A and 15B show matching points obtained by applying the conventional method and the present invention, respectively.

16A and 16B show face recognition results using the conventional method and the present invention, respectively.

As shown in FIGS. 14A to 16B, in the case of the present invention, the number of extracted feature points is increased by 30% or more on average, and thus, more matching points are obtained, thereby increasing face recognition rate. It can be seen that.

17A to 18B illustrate a face recognition result using the conventional method and the present invention for a face image with a lot of facial expression changes.

17A and 17B illustrate feature points extracted by applying the conventional method and the present invention to face images having a lot of facial expression changes, respectively.

18A and 18B show face recognition results of applying the conventional method and the present invention to face images having a lot of facial expression changes, respectively.

As shown in FIG. 17A to FIG. 18B, even in the case of the face image having a lot of facial expression change, more feature points can be extracted, thereby increasing the face recognition rate.

So far, the present invention has been described with reference to the embodiments. However, one of ordinary skill in the art will appreciate that the present invention can be implemented in a modified form without departing from the essential technical spirit of the present invention. Therefore, the disclosed embodiments should be considered in descriptive sense only and not for purposes of limitation. That is, the true technical scope of the present invention is shown in the appended claims, and all differences within the equivalent scope will be construed as being included in the present invention.

1 is a diagram illustrating feature points extracted by applying SIFT to two general images;

2 is a diagram illustrating feature points extracted by applying SIFT to two endoscope images;

FIG. 3 is a diagram illustrating feature points extracted by applying a SIFT having a concentration threshold of 0.01 and a curvature threshold of 10 for two endoscope images; FIG.

4 is a diagram illustrating feature points extracted by applying a SIFT set to a concentration threshold of 0.01 and a curvature threshold of 50 for two endoscope images;

5 is a block diagram showing an example of a face recognition system based on SIFT according to the present invention.

6 is a flowchart illustrating an example of a SIFT-based face recognition method according to the present invention.

7 is a block diagram showing an apparatus for extracting feature points based on SIFT according to the present invention.

8 is a flowchart illustrating a method of extracting feature points based on SIFT according to the present invention.

9 is a diagram illustrating a process of generating a DOG image using a Gaussian pyramid of an input image.

10 shows pixels of a DOG image compared to extract candidate pixels.

FIG. 11 is a diagram illustrating a process of extracting a directional distribution for 8 by 8 blocks and quantizing them in eight directions. FIG.

12A and 12B are diagrams illustrating feature points and matching points extracted by applying the present invention to a general image, respectively.

13A and 13B are diagrams illustrating feature points and matching points extracted by applying the present invention to an endoscope image, respectively.

14a to 18b is a view showing a face recognition result applying the conventional method and the present invention.

Claims (20)

An apparatus for extracting feature points of an input image based on a SIFT, A density threshold calculator for calculating a contrast threshold of a candidate pixel used for feature point extraction in response to an input image; And SIFT-based feature point extraction apparatus comprising a feature point descriptor generation unit for generating a keypoint descriptor by using the candidate pixel of the concentration threshold or more of all candidate pixels. The method of claim 1, The density threshold calculator may calculate a density value having a maximum number of pixels as the density threshold value through a histogram analysis of a difference of Gaussian (DOG) image generated from the input image. Feature point extraction device. The method of claim 1, The feature point descriptor generation unit generates the feature point descriptor based on feature point information including position, size, and orientation information of a neighboring pixel based on the candidate pixel above the density threshold. . The method of claim 1 The feature point extraction device, A DOG image generator for generating a DOG image using a Gaussian pyramid of the input image; And And a candidate pixel extracting unit extracting the entire candidate pixel by searching for the maximum or minimum point of the DOG image. A feature point based on a SIFT, calculating a contrast threshold of a candidate pixel used for feature point extraction corresponding to an input face image, and extracting a feature point using candidate pixels above the density threshold. Extraction unit; And And a face recognition unit configured to recognize the input face image by matching the feature point of the input face image with the feature point of the face image stored in the database. The method of claim 5, The feature point extraction unit, A density threshold calculator configured to calculate a density value having a maximum number of pixels as the density threshold value through a histogram analysis of a difference of Gaussian (DOG) image generated from the input face image; And And a feature point descriptor generator for generating a keypoint descriptor using candidate pixels above the density threshold among all candidate pixels. The method of claim 6, The feature point descriptor generation unit generates the feature descriptor based on feature point information including position, size, and orientation information with respect to neighboring pixels based on candidate pixels greater than or equal to the density threshold. . The method of claim 6 The feature point extraction unit, A DOG image generator for generating a DOG image by using a Gaussian pyramid of the input face image; And And a candidate pixel extracting unit extracting the entire candidate pixel by searching for the maximum or minimum point of the DOG image. The method of claim 5, And the face recognition unit matches feature points of the input face image and the stored face image through clustering by a Hough transform. The method of claim 5, And the face recognition unit recognizes the input face image in consideration of the number and distribution of matching feature points. A method for extracting feature points of an input image based on SIFT in a computer system, A density threshold calculating step of calculating a contrast threshold of a candidate pixel used for feature point extraction corresponding to the input image; And And a feature point descriptor generation step of generating a keypoint descriptor using candidate pixels above the density threshold among all candidate pixels. The method of claim 11, The concentration threshold calculating step may include calculating a density value having a maximum number of pixels as the density threshold value through a histogram analysis of a difference of Gaussian (DOG) image generated from the input image. SIFT based feature point extraction method. The method of claim 11, The feature point descriptor generation step may include generating the feature descriptor using the feature point information including position, size, and orientation information with respect to a neighboring pixel based on the candidate pixel above the density threshold. Feature point extraction method. The method of claim 11, The feature point extraction method, A DOG image generation step of generating a DOG image by using a Gaussian pyramid of the input image before the concentration threshold calculation step; And And extracting the candidate pixels by searching for the maximum or minimum points of the DOG image. A method for recognizing a face image based on a SIFT in a computer system, A feature point based on a SIFT, calculating a contrast threshold of a candidate pixel used for feature point extraction corresponding to an input face image, and extracting a feature point using candidate pixels above the density threshold. Extraction step; And And a face recognition step of recognizing the input face image by matching the feature point of the input face image with the feature point of the face image stored in the database. The method of claim 15, The feature point extraction step, A density threshold calculation step of calculating a density value having a maximum number of pixels as the density threshold value through a histogram analysis of a difference of Gaussian (DOG) image generated from the input face image; And And a feature point descriptor generation step of generating a keypoint descriptor using candidate pixels above the density threshold among all candidate pixels. The method of claim 16, The feature point descriptor generation step may include generating the feature descriptor using the feature point information including position, size, and orientation information with respect to a neighboring pixel based on the candidate pixel above the density threshold. Face recognition method. The method of claim 16 The feature point extraction step, A DOG image generation step of generating a DOG image using a Gaussian pyramid of the input face image before the concentration threshold value calculating step; And And a candidate pixel extracting step of extracting the entire candidate pixel by searching for the maximum or minimum point of the DOG image. The method of claim 15, The face recognition step may include matching feature points of the input face image and the stored face image through clustering by a Hough transform. 20. A recording medium having recorded thereon a program for executing a method according to any one of claims 11 to 19 on a computer system, the recording medium being readable by the computer system.

Images