A local patterns matching based human face recognition model and a oriental field based face detection model are presented. In the former, we detect the initial key features of a face image. Through local patterns matching, we get the true features that belongs to a human face. Then we can use them for face recognition. We use two methods to detect features, one is based on attention, the other is SUSAN. The former uses cooperation-competition mechanism to detect saliency feature points and uses multi-scale analysis to remove the effect of the texture. The latter is a new method for natural image processing, with very good veracity and stability. We use photopic map to get the local patterns of each feature point, and then represent the local patterns. We use KL transform to compress tile local patterns and memorize them. Then we match the pending local patterns with the memorized local patterns to decide which kind of patterns they fall into. The procedure of recognition is similar to the detection in our method. But we use two-grade matching, the matching of each local patterns and the matching of the space relations among these patterns, In the latter, we calculate the direction diagram and use the character of a human face in a direction diagram, according to some rules we present, we search the direction diagram to get the two feature parts of each face. Combining the two parts: we get each face in an image. The human face in a direction diagram is like a closed polygon(hexagon often). We consider two parts of it and get two kinds of feature areas. Some rules used are based on simple knowledge. It's a totally new and very quick method, the experiment results are quite exciting. The two methods both have some practicability.
修改评论