HMM-based architecture for face identification

Abstract This paper describes an approach to the problem of face identification which uses Hidden Markov Models (HMM) to represent the statistics of facial images. HMMs have previously been used with considerable success in speech recognition. Here we describe how two-dimensional face images can be converted into one-dimensional sequences to allow similar techniques to be applied. We investigate the factors that affect the choice of model type and model parameters. We show how a HMM can be used to automatically segment face images and extract features that can be used for identification. Successful results are obtained when facial expression, face details and lighting vary. Small head orientation changes are also tolerated. Experiments are described which assess the performance of the HMM-based approach and the results are compared with the well-known Eigenface method. For the given test set of 50 images, the HMM approach performs favourably. We conclude by summarizing the benefits of using HMMs in this area, and indicate future directions of work.