Robust face recognition using automatic pose clustering and pose estimation

The performance of most of the current face recognition systems drops significantly when there are variations in pose, illumination and expression. This paper mainly focuses on the pose problem in face recognition while considering the facial expressions and occlusions together. In this work, an automatic method for pose clustering and estimating the head pose from a single 2D face image and recognizing the face is presented. Gaussian Mixture Model is used for automatic pose clustering. Dense SIFT descriptors are extracted from image grid points in order to obtain a representation that is robust to pose, noise and illumination variations. PCA is used to reduce the dimension of the concatenated descriptor vector for efficient processing. In order to better approximate the head pose, a combination of Sparse approximation and Orthogonal Matching Pursuit (SAOMP) algorithm is used to infer a continuous mapping function from the image to the pose space. In addition, the proposed method is fully automatic. Extensive experimental results show that the proposed method achieves good recognition accuracy in spite of all uncontrolled conditions.