Partially occluded face recognition using subface hidden Markov models

An improved subface Hidden Markov model based on local texture with Discrete Cosine Transform (DCT) is proposed to recognize partially occluded faces. After LPQ (Local Phase Quantization) or LBP (Local Binary Pattern) features are extracted, the DCT is applied to generate the subface HMM´s observed vectors. All local subface HMMs of a face are then combined to a global HMM. Each partial occlusion is estimated by Haar features and is assigned with relevant weight automatically. The experiments on AR database show that local facial texture features transformed by DCT as HMM´s observed vectors are appropriate for partially occluded face recognition.