Based On The Statistical Characteristics Of Face Recognition Algorithms

As a combination of pattern recognition and image processing, face recognition is one of the most direct and friendly biometric authentication technology. With high research and application value, it is mainly used in the area of identity authentication and security monitoring. What the face recognition technology used is the human inherent characteristics, thus it has a high degree of security, reliability and validity. As one of the most easily accepted identity authentication methods, it has many advantages other recognition methods don’t have. For example, it can be collected easily and got naturally and directly, it both has the characteristic of universality and uniqueness, it can be used widely and so on.This dissertation mainly researched the methods of face image feature extraction and face recognition technology. The main work is summarized as follows:(1) Two image pre-processing methods that used in the face recognition systems were introduced:image noise elimination and image normalization. The image pre-processing methods were the key of improving the recognition rate. Its main effect was to standardize the collected images, thus to supply a preparation for the following training and recognition steps.(2) A face recognition method which based on the weighed K-L algorithm was adopted. This method increased the weigh of the eigenvector which was corresponding to the small eigenvalues for the effect of recognition results. The recognition effect of the K-L algorithm and weighed K-L algorithm with a series of simulation experiments were compared on different face image database and Gaussian white noise SNR. The results showed that, compared to the K-L algorithm, the weighed K-L algorithm had higher recognition rates and better anti-noise performances.(3)Two improved methods were adopted to solve the low recognition rate problem of the SVD algorithm:one is multi-scale segmentation of singular value vector based algorithm, another is K-L decomposition and SVD feature fusion based algorithm. These methods were experimented on different face image database and Gaussian white noise SNR. The results showed that, both of the two improved methods can improve the recognition rates and anti-noise performances a lot.(4) The initialization algorithm of the various parameters in the hidden Markov model(HMM) were introduced. In order to improve the recognition rate and reduce the dimension of the model observation sequences, K-L decomposition and SVD based method was used to extract the feature. The feature of the sub-image of the observation sequence extracted by this method was used as the observation sequence to train the model. Besides, the size of the sampling window and overlapping window on the effect of the recognition results were discussed. The optimal parameters were set by simulation experiments on different face image database. The recognition effect of the K-L+HMM method, the SVD+HMM method and DCT+HMM method were compared, and the result showed that, the K-L+HMM method and the SVD+HMM method can improve the recognition rate effectively.