Research On 3D-Face Data Acquisition In Binocular Vision

Face recognition is one of the most active research fields of authentication technology based on biological features, also a natural and friendly method of identity recognition. 3D face model has the crucial data of special structure (depth), so it has more rich information than 2D face image. It is hopeful to solve the problems in 2D face recognition caused by the variation of illumination, pose and expression. Therefore, 3D face recognition becomes a popular research. 3D-face data is the basis of 3D face reconstruction and recognition. However, how to obtain the 3D-face data is the important foundation of 3D face recognition.In this thesis, we do further research on obtaining the 3D-face data from the standard binocular face images based on the computer binocular vision technology. At first, the face regions are located by a face detection algorithm which is based on a skin color model. Then, its disparity image is work out by the datum-point-based stereo matching algorithm with an adaptive window. Furthermore, the 3D-face data can be obtained from this disparity image. Research works of this thesis are as follows:(1) In the research of face detection and location, illumination compensating is made for the color face image firstly. Secondly, a regional model of skin color in the YCbCr color space is built to segment the skin color in the image, according to the cluster trait of the skin color in color spaces. Finally, the face region is screened out from the skin areas. This method is simple, fast and insensitive to direction or poses, can detect faces from the complex backgrounds fleetly.(2) In the research of stereo matching, a reference-points-based stereo matching algorithm with an adaptive window is presented. It can achieve a relatively accurate disparity image by searching the adaptive matching window and growing the matching from the reliable-matched datum points. In the correction of mismatch, an improved zero-correction algorithm is proposed. It improves the accuracy of the disparity image effectively. Moreover, integral image technique is used in matching to simplify and speed up the computation. Before the matching, histogram equalization is used to reduce the gray values' diversities of the homonymy pixels in the two face-located gray images. It is helpful to improve the accuracy of matching.