Face recognition based on hidden markov model

Abstract

Computational efficiency is a matter of great concern in state-of-the-art face recognition systems. In this thesis, four-state hidden markov model for face recognition is proposed. Face images are divided into a sequence of overlapping blocks. An observation sequence containing coefficients of eigenvalues and eigenvectors of these blocks is used to train the model. Each subject is associated to a separate hidden markov model. Computational cost of proposed model is minimized by employing discrete wavelet transform on each image which reduces image redundancies while retaining informative features. Furthermore, principal component analysis is employed on image blocks that reduces blocks dimensions. A threshold eigenvalue is used to reject or accept test images. Accepted face images are classified based on the majority vote criteria using slightly different observation sequences of image features. Experimental results on Yale and ORL databases demonstrate that recognition accuracy of proposed model is comparable to the existing techniques. Test images of untrained databases are successfully rejected by the proposed model at a reduced computational cost. Keywords: Four State Hidden Markov Model, Discrete Wavelet Transform, Principal Component Analysis, Computational Complexity.