Face recognition using Pseudo-2D Ergodic HMM

The work presented in this paper describes a novel Pseudo-2D Ergodic Hidden Markov Model (EHMM) based architecture for automatic face recognition. The primary HMM of this model being ergodic in nature, gives the flexibility to switch between the states, contrary to conventional Pseudo-2D HMM, which follows a top-to-bottom approach. The new approach helps in better modeling the different variations of a human face. We present a segmental K-means algorithm for training the Pseudo-2D EHMM, thereby jointly optimizing the observation densities and the state transitions corresponding to different variations of the face. The performance of the proposed method is presented with Discrete Cosine Transform (DCT) and the DCT-mod2 feature sets for the Olivetti Research Laboratory (ORL) database. The better modeling capability of the proposed architecture along with the robustness of DCT-mod2 feature set to illumination direction changes, proves to be an excellent combination for automatic face recognition.