Research On HMM For Face Recognition

Personal identification system based on using the proper living creature characteristic of human body is the totally brand-new technique, is different from traditional methods and has the better safety, dependability and usefulness, so more and more people begin to think much of it. Person's face which has the directness, uniqueness and convenience et al is a unique living characteristic. Face recognition can be simply defined as follows: given a certain static image or dynamic video of scene, to detect and recognize one or more persons on the basis of prearranged face database, then to recognize who is(are) the person(s). It covers knowledge of many subjects, such as signal processing, intelligence control, pattern recognition, machine vision et al, has very high theoretic values, and becomes a very active research topic in computer vision and computer pattern recognition. With the development of intelligent information and processing, face recognition will be broadly applied in law, business, security system, and so on.The main work primarily includes:1. On the basis of several kinds of face recognition methods, HMM is mainly studied. A new algorithm which focuses on the use of Integral Projection Functions and Hidden Markov Models for face recognition is presented. First, the Integral Projection values of face images are computed. Second, these Integral Projection values which are translated into one- dimension vector sequences are trained and recognized by HMM. The result of experiment on ORL database shows the improved recognition rate.2. A new algorithm which focuses on the use of Fisher Linear Discriminate Analysis, Principal Analysis in the Complex Space and Hidden Markov Models for face recognition is presented. First, the different images are translated into one-dimension vector sequences with the same mean and variance. Second, FLDA is used to get the features of the images and complex vector space, CPCA is applied to get the new features, then these new features are trained by HMMs. Finally, an optimized HMM is obtained. Compared with other face recognition algorithms on the ORL face database, this method can get better recognition rate.