Face Recognition

Face recognition is a technology using computer to analyze the human face image and extract the features for recognizing the identity of the target. The research in face recognition has been to a hot spot of the category of artificial intelligence and pattern recognition. Face recognition has three steps: face detection, feature extraction and recognition. The face recognition can be applied to: intelligent entrance guard system, computer security, machine vision, medical diagnostic system and 3-D animation etc. As a research domain with wide applications, face recognition absorbs lots of researchers. Some famous research institutes also list it as important research direction and achieve great accomplishments. Although the face recognition is mature under the ideally situation, many problems need to be solved to develop a robust, large-scale and speedy face recognition system.There are a lot of algorithms about the face recognition. In recent twenty years, mangy research institutes and researchers proposed their methods. But according to their different academic backgrounds and applications, these algorithms are different widely. ASM (Active Shape Model) and AAM (Active Appearance Model) are two efficient algorithm for feature extraction. They are both parametric models based on statistics and widely used for the location of the facial key points, medicinal image processing, image registration and tracking etc.In this paper, we research in the modeling process of the ASM and AAM, and compare them. Then we analyze the fitting algorithm of the AAM deeply. At last, we proposed our improvement both on ASM and AAM. We introduce our works concisely as follows:1,Reproduce the modeling process of the traditional ASM and AAM. We apply the PCA onto the training sets to build the parametric ASM and AAM models by extracting the shape information by handy location and the global texture by warping. Then we compare their performance by the experimental results.2,According to the researches on the alignment of the training sets in traditional ASM, we find that the Procrustes algorithm is an iterate method. It will be low efficient when the training sets is very big. So we proposed an novel algorithm based on the geometric transform to align the training sets to an unified framework. The experimental results indicate that the new algorithm reduce the computation times remarkably.3,We compare the linear algorithm and the inverse composition algorithm used for the fitting of AAM and analyze the reasons which lead to different fitting results. According to the experimental results, we acclaim that the inverse composition algorithm is accurate.4,Based on the fitting process of the AAM, we analyze the influences cased by the global geometric transform. The fitting results will be better if we add the global geometric transform to the inverse composition algorithm. At last we add the shear transform to the algorithm besides the basic global geometric transform, which leads to a better fitting result. The experimental results prove that the algorithm is efficient.