The Research Of2D Face Recognition Combined With3D Model

As an important biological recognition technology, face recognition has the advantages of non-touching, safety and convenience, and has been used widely. Although the study of face recognition has undergone decades and most current face recognition systems work well under constrained conditions, the performance of most of these systems degrades rapidly when they are carried on under uncontrolled conditions. The existing recognition methods have some defects, which make the research of face recognition still have very high research and practical significance.In order to deal with such bottleneck problems as complex illumination and the change of face pose in the research for two-dimension face recognition:, and to solve the complex computation and the widely application difficulty in three dimensional face recognition, this paper proposes a2d+3d face recognition technology. There are three core ideas in the method:using illumination rendering and multiple perspective projection transformation in3d model to generate the training samples; face pose estimation based on hierarchical SVM; matching the training samples that have the same estimated pose. The2d+3d face recognition can be divided into several stages, and the main contents are listed as follows:1) According to the symmetry of the face, this paper uses the front and the left side of the face detector which are trained by the Adaboost algorithm to detect the multi-view face.The detection rates esting on LFW and FacePix are86.6%and89.4%respectively.2) This paper uses a model based on a mixture of trees with a shared pool of landmarks to obtain facial landmarks on contour and cut out the face region by the contour landmarks. Then we use the gamma nonlinearity, Difference of Gaussians filter, equalization and a series of operations to reduce the influence of illumination.3) We obtain a hierarchical SVM for face pose estimation through training the projection images of BJUT-3D models. The accuracies of pose estimation testing on FERET images and projection images of BJUT-3D models are82.75%and90.73%respectively.4) With making best use of3D model and texture information, the training samples are obtained according to the illumination rendering and multi-view perspective projection transformation. Then the obtained samples are used to recognize faces with complex illumination and posture in real life by using weighted chi square distance. Through test on BJUT, the proposed mothed can achieve the recognition rate reaching91%.