WO2008066217A1

WO2008066217A1 - Face recognition method by image enhancement

Info

Publication number: WO2008066217A1
Application number: PCT/KR2007/000154
Authority: WO
Inventors: Donghoon Jeon; Sungha Shin; Jonghwa Cheon
Original assignee: Firstec Co., Ltd.
Priority date: 2006-11-30
Filing date: 2007-01-10
Publication date: 2008-06-05
Also published as: KR20080049206A; KR100843513B1

Abstract

The present invention relates to a face recognition method by image enhancement and, more particularly, to a face recognition method by image enhancement which extracts a face image of a suspect from a moving image or a still image that contains a scene of a crime, processes the extracted face image, compares the processed face image with faces of criminals stored in a database having face information of criminals stored therein so as to identify the suspect. The face recognition method comprises a face extracting step of extracting a face image from an image that is obtained from a CCD camera and contains the face of a suspect, an image correcting step of correcting the face image, a characteristic extracting step of extracting characteristic of the face, and an identification step of comparing the face image with pictures stored in a database to identify the suspect, wherein the image correcting step enlarges the face image to a predetermined size in order to compare the face image with the pictures stored in the database, calculates pixel values of pixels of the enlarged face image using an interpolation method according to the position of the enlarged face image and contour information, and processes the calculated pixel values according to a statistical analysis.

Description

Description FACE RECOGNITION METHOD BY IMAGE ENHANCEMENT

Technical Field

[1] The present invention relates to a face recognition method by image enhancement and, more particularly, to a face recognition method by image enhancement which extracts a face image of a suspect from a moving image or a still image that contains a scene of a crime, processes the extracted face image, compares the processed face image with faces of criminals stored in a database having face information of criminals stored therein so as to identify the suspect.

[2]

Background Art

[3] Recently, crimes such as bank or shop robbery and extortion of money from drunken men or the old and the weak occur frequently. Furthermore, a crime is frequently committed in which cash is drawn out from an automated teller machine (ATM) using a card extorted from a person.

[4] In order to prevent the aforementioned crimes and acquire information on suspects,

CCTV cameras are installed in places having high possibility of crime. Moreover, cellular phones attached with cameras are widely spread, and thus pictures of crime scenes, captured by witnesses, are provided to the police in many cases.

[5] However, a camera with a high resolution is required in order to store images obtained from CCTV cameras or cameras attached to cellular phones with distinctness. Furthermore, CCTV cameras are outworn equipment, so that they have a low resolution, and it is difficult to identify suspects because most criminals hide their faces. Moreover, even if the face of a suspect is clearly captured by a camera, the police have no choice but to spread a search for the suspect and receive a report from an information provider in order to identify the suspect.

[6] In order to solve the aforementioned problems, attempts to use face recognition systems installed in various buildings or offices have been made recently. Most face recognition methods are being studied which enlarge a face image, extract characteristics of a face and compare the face image with pictures of criminals stored in a database.

[7] Japanese Patent Laid-Open Publication No. Hei05-266173 discloses a face recognition technique that extracts a face image, removes the influence of lighting using a homomorphic filter, and generates rotation, magnification/reduction of the face image according to a recursive call second-order moment segmentation means to represent a face using a characteristic vector. That is, the positions of eyes and a mouse in a face region are determined, the face is segmented with a line connecting the two eyes and a line that is perpendicular to the line and passes through a nose, and second- order moment segmentation is limited to the face region to remove noise that affects the characteristic vector so as to recognize the face.

[8] Japanese Patent Laid-Open Publication No. Hei07-302327 discloses a technique that captures data of a face in various directions, stores the data, and compares the data with stored images to detect an image having highest similarity.

[9] Furthermore, Korean Patent Laid-Open Publication No. 1999-50271 discloses a face recognition method that extracts face images of suspects and identifies a most similar suspect. This face recognition method includes a face extraction step of extracting a face image from a gray picture or a color picture, an image correction step of correcting the face image, a characteristic extraction step of extracting characteristic of a face, and an identification step of comparing the face image with pictures stored in a database to identify a suspect having the face image.

[10] The face extraction step uses a method using the motion of eyes (disclosed in

Korean Patent No. 361497), a method using grouped face images and a mesh type search region (disclosed in Korean Patent No. 338807), a method of detecting a face region with a face color using edge/color information (disclosed in Korean Patent No. 427181), or a method using an adaptive boosting algorithm based on rectangular characteristic of a face (disclosed in Korean Patent No. 621883).

[11] The characteristic extraction step uses face characteristics disclosed in Korean

Patent Laid-Open Publication No. 1999-50271, and a hierarchical graph matching method using a flexible grid and a principal component analysis method are known as a characteristic extraction technique.

[12] The identification step that compares a face image with pictures stored in a predetermined database to identify a suspect having the face uses a method of determining a hyperplane through a support vector machine and learns a face recognition method to recognize a face (disclosed in Korean Patent No. 456619, No. 571826 and No. 608595), a hierarchical principal component analysis method (disclosed in Korean Patent No. 571800) and so on.

[13] However, while the face extraction step, the characteristic extraction step and the identification step are all related to automation for identification, images (including still images of moving images) acquired by CCD cameras from a criminal scene show a very small suspect's face in many cases. In this case, it is required to magnify and correct the images in order to obtain distinct characteristics of a face and compare the face with pictures stored in the database.

[14] The images are mostly magnified through an image processing program such as photoshop. In order to automate magnification of images, the following technique is used.

[15] In general, such an image captured by a camera, as described above, is a digital image, and thus the image is composed of pixels to which RGB values or contrast values are respectively allocated (hereinafter, information allocated to each pixel, such as an RGB value or a contrast value, is referred to as 'pixel value'). To magnify the image, a technique that repeatedly uses a pixel value of a pixel for neighboring pixels according to a magnification is generally used. More specifically, pixel values are repeated twice and recorded to magnify the original image twice, as illustrated in FIG. 1. FIG. 2(b) illustrates an image of 220x220 pixels, which is obtained by magnifying an image of 55x55 pixels illustrated in FIG. 2(a) four times. Referring to FIG. 2(b), the enlarged image of 220x220 pixels shows repetition effect compared to the image illustrated in FIG. 2(a).

[16] In addition, linear interpolation that interpolates the pixel value of a corresponding pixel using an average value of pixel values of neighboring pixels larger than magnification is used to magnify an image. In this case, an edge region is smoothed to generate blurring, as illustrated in FIG. 2(c), and thus the face extraction step and the characteristic extraction step cannot produce satisfactory results. Furthermore, an image can be transformed to frequency components, and then Fourier-transformed to be magnified. In this case, noise is added to the image in a low frequency region while edge characteristic is maintained.

[17] The aforementioned conventional method is difficult to obtain satisfactory results in the identification step because serious low-frequency noise is generated in an edge region or blurring occurs in the characteristic extraction step. Furthermore, even though a face image is corrected, the best combination of various methods known as the face extraction step, the characteristic extraction step and the face recognition step should be found in a face recognition field.

[18]

Disclosure of Invention Technical Problem

[19] Accordingly, the present invention has been made to solve the above-mentioned problems occurring in the conventional art, and a primary object of the present invention is to provide a face recognition method capable of efficiently detecting a face region, magnifying the detected face region with distinctness, and effectively comparing the detected face region with characteristic parts of a face. Technical Solution

[20] To accomplish the object of the present invention, there is provided a face recognition method by image enhancement comprising: a face extracting step of extracting a face image from an image that is obtained from a CCD camera and contains the face of a suspect; an image correcting step of correcting the face image; a characteristic extracting step of extracting characteristic of the face; and an identification step of comparing the face image with pictures stored in a database to identify the suspect, wherein the image correcting step enlarges the face image to a predetermined size in order to compare the face image with the pictures stored in the database, calculates pixel values, i.e., contrast or color information of pixels of the enlarged face image using an interpolation method according to the position of the enlarged face image and contour information, and processes the calculated pixel values according to a statistical analysis.

[21] A low-frequency noise is removed by means of a wavelet filter for the pixel values processed according to the statistical analysis.

[22] The face extracting step uses an AdaBoost method, the characteristic extracting step uses an HGM method, and the identification step uses an SVM method.

Advantageous Effects

[23] According to the present invention, a low-resolution image is magnified without distorting the characteristic of the low-resolution image when the low-resolution image is converted to a high-resolution image, and thus a face region can be detected more efficiently, the detected face region can be magnified more distinctly, and the face region can be compared with face characteristic parts more effectively.

[24]

Brief Description of the Drawings

[25] Further objects and advantages of the invention can be more fully understood from the following detailed description taken in conjunction with the accompanying drawings, in which:

[26] FIG. 1 illustrates pixel values of an image when the image is magnified using a conventional method;

[27] FIG. 2 illustrates an image magnified according to a conventional image magnification technique;

[28] FIG. 3 illustrates edge/contour extraction according to the present invention;

[29] FIG. 4 illustrates interpolation between contours according to the present invention;

[30] FIG. 5 illustrates image enlargement and correction according to the present invention; and

[31] FIG. 6 illustrates an image magnification and correction result according to the present invention.

[32]

Mode for the Invention [33] In an embodiment of the present invention, a low-resolution image of the face of a suspect, obtained from a CCD camera, is explained.

[34] Edges are extracted from the low-resolution image obtained from the CCD camera, as illustrated in FIG. 3 (a). An edge is designated as pixels having contrast values exceeding a predetermined threshold value and has position values (ui, vi) of a series of pixels. A contour function f(ui, vi) that represents contours is obtained according to edges in a predetermined range among the extracted edges of the low-resolution image.

[35] The contour function f(ui, vi) can be a linear function or a nonlinear function. The contour function f(ui, vi) includes multiple functions fk(ui, vi) respectively corresponding to multiple sections of the edges of the image such that the contour function f(ui, vi) satisfactorily reflects the position values (ui, vi). Contours and edges obtained according to the contour function are illustrated in FIG. 3(c).

[36] A contours represented according to the contour function is a continuous model represented in the form of a segment of a line while a pixel is an independent discrete model, and thus the characteristic of the contour is maintained even when the contour is enlarged, as illustrated in FIG. 3(c). That is, even when the low-resolution image is converted to a high-resolution image, edge characteristic of the low-resolution image is maintained.

[37] For example, when a low-resolution image of 55x55 pixels, as illustrated in FIG.

2(a), is converted to a high-resolution image of 220x220, as illustrated in FIG. 2(b), repetition effect is produced, and pixels producing the repetition effect form a thick edge. This thick edge makes representation of a distinct contour difficult. However, the characteristic of a contour is maintained in the high-resolution image even though the contour is obtained from the low-resolution image, and thus edge characteristic is maintained.

[38] Pixel values of pixels constituting an edge in the low-resolution image can be maintained even in the high-resolution image according to the contour function and the width of the edge is also maintained in the high-resolution image. Pixel values of pixels constituting an edge in the high-resolution image maintain the original pixel values and other pixel values copy corresponding pixel values of the low-resolution image multiple times corresponding to magnification in such a manner that FIG. 2(a) is converted to FIG. 2(b) to form the high-resolution image.

[39] Interpolation according to successive linear analysis or nonlinear analysis is performed on pixel values existing between edges constructing the contour function and neighboring edges to calculate pixel values of pixels between the contours. This is explained in more detail.

[40] FIG. 4(a) illustrates a part of an image including multiple edges. A grid indicated by a solid line (that is, a grid formed by first and second rows and first and second columns) represents a pixel of a row-resolution image and a grid indicated by a solid line and a dotted line (that is, a grid formed by the first row and the first column) represents a pixel of a magnified high-resolution image. Pixel values of pixels corresponding to contours maintain the original pixel values and pixel values of enlarged and newly formed pixels are calculated. Referring to FIG. 4(b), pixel values of pixels existing between pixel values of pixels corresponding to contours are calculated according to interpolation. Here, multiple pixel values are obtained for pixels existing between the contours. For example, for the pixel of a fifth row and a seventh column, seven pixel values are obtained even though calculation is performed on only two pixels indicated in FIG. 4(b). A pixel value with the highest frequency among multiple pixel values allocated to pixels, obtained through calculation according to interpolation, is determined through statistical analysis or an average value of the multiple pixel values is determined. When the image illustrated in FIG. 2(a) is processed through the image processing technique according to the present invention, an image illustrated in FIG. 6 is obtained. Referring to FIG. 6, an excellent enlarged image having distinct edge characteristic and no low-frequency noise can be obtained.

[41] Accordingly, image correction that produces satisfactory edge characteristic and removes repetition effect is achieved. An image has low-frequency noise when the contour function is designated using a considerably small number of edge position values. In this case, it is preferable to remove the low-frequency noise using a wavelet filter.

[42] When a face region is extracted according to adaboost method from an image corrected by the above-described method, face characteristic is extracted using HGM method, and a face is recognized according to SVM method, a face detection success rate is increased to 95% from 82% and a recognition rate of higher than 88%, 95% and 95% is obtained even though lighting/expression/pose are changed.

[43]

Industrial Applicability

[44] The present invention relates to a face recognition method and, more particularly, to a face recognition method which extracts a face image of a suspect from a moving image or a still image that contains a scene of a crime, processes the extracted face image, compares the processed face image with faces of criminals stored in a database having face information of criminals stored therein so as to identify the suspect. According to the present invention, a high recognition rate can be secured.

Claims

[1] A face recognition method by image enhancement comprising: a face extracting step of extracting a face image from an image containing the face of a suspect; an image correcting step of correcting the face image; a characteristic extracting step of extracting characteristic of the face; and an identification step of comparing the face image with pictures stored in a database to identify the suspect, wherein the image correcting step enlarges the face image to a predetermined size in order to compare the face image with the pictures stored in the database, calculates pixel values, i.e., contrast or color information of pixels of the enlarged face image using an interpolation method according to the position of the enlarged face image and contour information, and processes the calculated pixel values according to a statistical analysis.

[2] The face recognition method by image enhancement according to claim 1, wherein a low-frequency noise is removed by means of a wavelet filter for the pixel values processed according to the statistical analysis.

[3] The face recognition method by image enhancement according to claim 1 or 2, wherein the face extracting step uses an AdaBoost method, the characteristic extracting step uses an HGM method, and the identification step uses an SVM method.