Face Recognition Technology Based On Singular Value Decomposition And Embeded Hidden Markov Model

Machine recognition of Human face is currently one of the most active researches and the most challenging problems. At first, the face images are obtained by different way or in different condition, so they have substantial difference in quality, geometry, illumination, etc. But the most essential reason is that face is a kind of non-rigid object that has highly similarity. Different person's faces have similar shape and structure, and one person's face has his different state in different conditions. In the past decade, many research groups make great efforts on it and a series of successes have made personal identification appear not only technically feasible but also economically practical. However, no perfect solution can accomplish this task under the non-constraint condition.The methods of face recognition involve four major fields: based onappearance feature, based on template, based on algebra feature, and based on machine learning.It is of particular interest in a wide variety of applications. In fact, face recognition technology has two kinds of application: recognition/ identification/ match and verification/ authentification/ surveillance. A typical example about the former one is applications in law enforcement for mug-shot identification. The latter application is also broad, such as verification for personal identification, gateways to limited access areas, authentification for ATM and family security, etc. In brief, one is concluding who is he/she and the other is deciding is he/she somebody.The followings are the main research contents in my paper:First, we use Local Singular Value Decomposition method to subtract the facial features .As one improved method, the way it works on according on following steps: first, facial images are divided to several area, and calculating the singular value vectors of each area, which are composed to global singular value vector sequence of hole facial image. Through the test, it is concluded that the main feature of person's face are included in the several anterior no zone vectors, with which big differences exist, so, we only select a few anterior vectors no zone in application, it remain the useful information and reduced the calculation quantity and the restore quantity.Second, the approach for facial recognition based on Embed Hidden Markov Model is the emphases of the paper, it developed the traditional one dimension Hidden Markov Model, so called P2D Hidden Markov Model, the approach meliorated the problem of plotting statuses so simple in one dimension HMM, and, avoided the high complexity in 2D HMM, it is just a better model structure to describe the feature of face. In the research procedure, we ensured partial model parameters of EHMM classifier: the size of window for selecting the observation vector and the number of mixture Gauss, which making the recognition ratio and the time of training and recognition perfect balance. Our experiment achieved the facial recognition algorithm based on HMM and EHMM each and tested in face database ORL and YALE, and, validated the advantage of EHMM through comparing the recognition time, training time and recognition ratio in above algorithms and face database.Third, since the automatic facial recognition compose with facial detection, facial image preprocess, subtracting the facial feature, classifier design, and decision part, after subtracting the facial feature, designing classifier mentioned above we have following works to do:1. Facing factual applications, the part of detection used AdaBoost fast face detection algorithm, which is explored to train and classify the cascade system, The principle of boosting is that a series of weak classifier, whose performance is only better than random suppose, is combined into a perfect ensemble strong classifier. An algorithm based on AdaBoost trains weak classifiers. Adaboost is an iterative method for obtaining an ensemble of weak classifiers by evolving a distribution of weights. Three simple types of the rectangle features are used to represent the feature of human face. And a perceptron, whose structure is very simple, is adopted to reduce the complexity of compute. The process of training is to select individual and valid features from a very large feature set. Each classifier node is designed to reject a potion of non face regions. Most image regions are rejected quickly, resulting in very fast face detection performance, and it meet the request of real time detection with 20 frame a second.2. A new algorithm for illumination compensation based on wavelet is applied. The method reduces and gets rid of the negative influence of the fluctuation of illumination. The method is applied to improving variable illumination of human facial image so that there are less restricted conditions in a face recognition system. Several empirical tests are given to demonstrate the effectiveness of our method. The method can be applied to a real system. And it can also improve the system robustness and adaptability under varying illumination conditions.Finally, Based on the above theory and experiment, software has been...