1D-HMM for face verification: model optimization using improved algorithm and intelligent selection of training images

In this paper, we present an optimized version of 1D-HMM for real-time face verification. DCT coefficients of face images are used as observation vectors in HMM states. Three modifications have been proposed to improve the overall performance of the approach: (1) replacing Baum-Welch algorithm with a clustering algorithm, (2) adding a clustering performance measure to the clustering algorithm and (3) selecting an intelligent training set among available images in data set. Despite its lower computational complexity, this approach shows better verification performance compared with other 1D-HMM methods. The proposed algorithm has been successfully tested on the well-known ORL face data set, exhibiting an accuracy of 96%. This is more than 10% higher than the verification results of the classical 1D-HMMs and is comparable with the results obtained with the 2D-HMMs, which is much more complex than the 1D-HMM.