3D Face Recognition Based On Collaborative Representation Classification

The current face recognition is mainly based on two-dimensional intensity image. It is simple, fast and effective, but the performance of recognition rate will decrease sharply once the environment of tester affected by the changes of illumination, face expression, gesture and shelter, which greatly limit the further development and application of face recognition. Development of three-dimensional imaging technology makes it possible to acquire 3D depth map data which is not sensitive to light changes and contains the inherent feature.3D face recognition is becoming more popular and attracting more attention. For overcoming the effect brought by the varieties of illumination, expression or gesture, the thesis studies 3D face recognition and proposes two face recognition methods based on collaborative representation classification algorithm.Firstly, the thesis outlines the background and significance of face recognition, introduces the development and the situation, and describes the basic technologies such as facial feature selection, feature extraction, classification and so on. As the foundation of proposed algorithm, sparse representation classification and collaborative representation classification is studied roundly in the next part.Secondly, insufficient training information will lead to a poor recognition result when the training set of each sample is small. To overcome this problem, we combine 2D face intensity image with 3D face depth map in complex domain, aim to increase training information, and define a similarity measure on collaborative representation algorithm to classify testing faces. The experiment results on Correlation Image Sensor database and Texas face database demonstrate the effectiveness of proposed method.Lastly, we propose a kernel collaborative representation classification based on Gabor feature. In this method, by extracting the Gabor feature of 3D face depth map, we retain more discriminative information in face structure so as to decrease the effect of illumination, facial expression and pose variants, and then fuse kernel learning in classification. The experiments on Kinect dataset and Texas face database show that this method improves recognition performance under the variants of illumination, expression and gesture.