Face Recognition Algorithm Based On Quasi-Binary Image

Face, as the most important biological characteristics, plays a crucial role in people's daily interactions, and it's one of the topics as the most challenging area in pattern recognition and machine vision. Face recognition system mainly includes face detection, facial feature location, as well as face recognition of three parts, meanwhile, the human eyes localization plays a key role in the face detection and recognition system. In this paper, picture based on quasi-binary with minimal gray values is researched. Projection functions, angle between space vectors and improved HAUSDORFF distance are concerned as research tools to study eyes location and face recognition, respectively, and a new localization algorithm and two recognition models are given. Major work are as follows:(1) A algorithm for eyes localization based on twice binary images and twice vertical projections is described in this paper. Reconstructed quasi-binary image by wavelet and new singulars builds a better platform for future eyes position work. Experimental result indicates that the proposed method is easy, feasible, and fairly precise for eye localization.(2) A new integration algorithm based on Singular Value Decomposition and angle between space vectors for face recognition is mentioned. Quasi-binary image is the same recognition basic, but the difference is that different singular vectors obtained by the whole and partial SVD methods are taken as vectors to be recognized. At last, angle between space vectors is used to identify persons. Experimental result indicates that both methods get a satisfied recognition rate, meanwhile, weighted partial SVD algorithm is more effective than the whole SVD algorithm.(3) An algorithm based on quasi-binary image and the improved HAUSDORFF distance for face recognition is given. The transformed quasi-binary image is to be identified. Result shows that it's more effective than the previous methods. At the same time, It is worth noting that the above two methods eliminate the need for training images, increasing the speed of operation, it also achieves a good recognition results at the same time.