Research And Application On EHMM-Based Face Recognition

Face recognition is an area of research spanning several disciplines, including imageprocessing, pattern recognition and computer vision, etc. Face detection andrecognition is one of the most basic and significant steps in the interaction applicationdomain such as intelligent identification, video-surveillance, lip-reading and facialexpression extraction. For the extraction procedure of facial feature is moreconvenient than any other biological features, scarcely requires the user’s complicityand is user-friendly, face recognition is regarded as the biological feature recognitiontechnology with great promise in the future.. A complete face recognition system iscomposed of preliminary image processing, face detection and location, facerecognition and matching, and specific applications, etc. This paper discusses thebasic face detection and recognition methods and proposes some improvements, themain work is as follow:(1)In the classifier-based face detection, multilayer perceptron is selected as thetopology of artificial neural network to meet the need of better performance inseparating capacity. For that gray level does not represent the robust feature, whichbrings non-deterministic to two-class classification problem, this paper proposes amethod using Gabor filter as the input vector, the experimental results show that ispractical.(2)Considering that sampling size, overlap and DCT coefficients numbers havean impact on the efficiency of hidden Markov model-based face recognition, theimproved method increases the training speed and recognition rate of the face modelthrough optimizing the combinations of input parameters. At the stage of imagematching, the extended Viterbi algorithm is estimated by a threshold to solve thetime-consuming problem caused by the discriminability-lacking among the models,which has almost equal accuracy, with reduced time for recognition.(3)An embedded hidden Markov model is an extension of1D-HMM in order todeal with2D-data such as image. In the capacity of describing and differentiating,EHMM has better performance than HMM, and this promotion also lead to theincrement in computational complexity, so a simplified EM algorithm is proposed in this paper, the algorithm operates in super state and sub state direction respectively todecrease the complexity of EM.