Occlusion expression recognition based on non-convex low-rank double dictionaries and occlusion error model

Abstract

Aiming at effectively correcting recognition error caused by partial facial occlusion, and simultaneously reducing the influence of facial individual′s difference on expression recognition, a novel method based on double dictionaries got by non-convex low-rank decomposition and occlusion error model is proposed in this paper. There are two steps in the training stage: 1° we use logarithmic determinant function for rank minimization problem of facial feature matrix, whose singular value fluctuates rather largely among different dimensions and causes that normal nuclear norm function cannot effectively estimate the minimal rank of the matrix. 2° non-convex low-rank decomposition algorithm is used to learn double dictionaries for expression features and individual identity features, respectively, and then an intra-class related dictionary and an inter-class structured dictionary are obtained. Three steps are taken in test stage: 1° we establish an occlusion error model to represent bias information of an occluded test sample, which is represented in a single matrix and added into iteration reconstructing process of the occluded test sample, based on double dictionaries obtained in the training stage. 2° The occluded test sample is decomposed into three sub-images, including expression feature, identity feature and occlusion error. 3° The occluded test sample is represented by the decomposed expression feature sub-image and then is classified into corresponding expression category. The following experiments in CK+ and KDEF expression databases show that, the proposed method is robust to the recognition of random occlusion facial expression images.