Face Recognition Algorithm Based On Compressed Sensing

Terrorist attacks frequently occurred since the twenty-first century. Public safety issues become increasingly prominent. Identification recognition technology plays a more and more important role in security domain. Traditional identification technology,such as password, magcard and code, is easy to lose, forget and steal, so people’s attention transfer to biometric identification technology. As one of the outstanding representatives, face recognition has been widely applied in security monitoring, identity authentication, entry and exit administration, crime picture retrieval and so on because of its advantages, for example unconstraint and lowcost. Simultaneously, face recognition is a multi-disciplinary crossing technology. Its improvement can promote the development of machine learning, computer vision and related areas. Therefore it has important research significance.The existing face recognition algorithms achieve good results under constraint environment,however, they still have many problems in a real world application. In2009, Wright et al present a recognition algorithm based on compressive sensing: sparse representation classification algorithm, which can effectively solve the problems of light and occlusion. However, SRC algorithm is too sensitive to the error of alignment. When the face’s pose or expression changes, the recognition rate of SRC drops rapidly, so it cannot be used in the real environment. On the other hand, SRC algorithm calculates the coefficient of sparse representation through minimizing the L1norm of the coefficient. This process’s computation is very large, cannot satisfy the fundamental real-time requirement, which further reduce the practicability of the algorithm. The main research content and innovation point of this paper are as follows:1. We propose a face feature extraction method which is robust to pose variant based on facial feature detection. Mark the pose of each training sample, then train a forest which can detect the pose of input face picture. Take out subset of the training data according to the pose, train conditional forest with them to detect the facial feature points of the face. Then estimate the affine transformation matrix which can rotate the face to horizontal position via the facial feature point. Do PCA transform on the whole face as the global feature. Extract LBP feature on the image sub-block around the facial feature point as local feature. Concatenate the two features and then do dimensionality reduction to construct the final descriptor which is robust to pose variant. Experiments prove that this descriptor improve the robust to pose and expression variant of SRC algorithm.2. We propose a real-time face recognition algorithm based on Fisher discrimination dictionary learning. First design a dictionary learning objective function. This function guarantee a face can be best represented by the samples of its own, cannot be represented by the samples from others. Simultaneously Fisher discrimination make the coefficients of the registered sample set have a small within-class scatter but big between-class scatter. Then calculate the coding coefficients though a simple regularized least square on this dictionary. Finally, we identify the face with the reconstruction error, coefficient’s Euclidean distance between test sample and registration samples, test sample’s L2-norm of coefficient correspond to each category.