Research On Faces Recognition Under Partial Occlusion Based On Sparse Representation And Nonnegative Matrix Factorization

With the strongly increasing of applications requirement for face recognition technics, unconstrained faces recognition has received widely attentions and researches. Partial occlu-sion is one of the main problems in the unconstrained faces recognition. Partial occlusion, which leads the face information to be incomplete, severely affects the feature extraction and recognition and has not been solved efficiently. For the further extensive application, it is an important project to develop new recognition methods which are robust to occlusion.Up to now, there are still some drawbacks for the researches on faces recognition under partial occlusion:(1) Most approaches can only deal with small partial occlusion and fail to the large contiguous occlusion, such as scarf.(2) The features extracted by nonnegative matrix factorization are not robust to occlusion because the geometric structure of the faces manifold, the long tailed property of reconstruction error distribution and supervised infor-mation are not considered simultaneously.(3) Standard sparse representation assumes the coefficients are independently each other, and does not consider geometric structure of the faces manifold in the model.Based sparse representation and nonnegative matrix factorization, this paper thorough-ly studies faces recognition under partial occlusion by utilizing machine learning methods from following three levels:(1) First, we consider the scenario that the supervised information and occlusion train-ing samples are available simultaneously. For large contiguous occlusion, we propose a robust face recognition algorithm based on occlusion dictionary learning. We solve this problem from following three aspects:1) We propose the occlusion dictionary to model the occlusion pattern. Each atom of the occlusion dictionary presents a kind of occlusion such as sunglasses or scarf, the occlusion in the test sample is sparsely represented by the atoms from the occlusion dictionary.2) We present the occlusion dictionary learning model based on mutual coherence minimum. The non-occluded face is sparsely represented by the training sample dictionary, while the occlusion is sparsely represented by the atoms from occlusion dictionary. By minimizing the mutual coherence between those two dictionaries, the occlusion parts can be efficiently separated, then the recognition can be conducted on the recovered faces.3) We propose a fast algorithm to learn the occlusion dictionary and prove the convergence. In order to accelerate convergence, only one atom is updated in each step, and the updated atoms are used in the next atom updating. Experimental results show that the atoms from the learned occlusion are more like the occlusion, e.g., sunglasses, scarf or extreme illumination variation, the recognition rate also shows that the proposed method is more robust than other methods for the large contiguous occlusion.(2) Then, we consider the scenario that the supervised information is available, but there are no occlusion training samples. For the sensitiveness to occlusion of nonnegative matrix factorization, a robust nonnegative patch alignment framework is proposed based on local geometric structure and supervised information. Correntropy induced metric is adopted to measure the reconstruction errors based on the long tailed property of the recon-struction errors distribution. The more large the error for each pixel, the more large weight is assigned. The weight is learned adaptively according to the error during the iteration process. For the non-convexity of the model, the problem is formulated as half-quadratic optimization problem via auxiliary variables. The concise update scheme is obtained via iterative optimization schemes and the convergence is proved. As application of this frame-work, locality preserving nonnegative patch alignment (LP-RNA) and sparsity preserving nonnegative patch alignment (SP-RNA) are proposed. In the LP-RNA, a locally sparse graph is presented to characterize the relation of occluded faces. First a suitable large set is selected by k-nearest neighbors for each sample, and then sparse representation is im-plemented in this neighbor. This way combines the locality of k-nearest neighbor and the robustness of sparse representation. In the SP-RNA, we utilize reconstruction coefficients to characterize the local geometric and use weighted distances to characterize separability of classes, then we construct the part optimization model. The test results on synthetic occlu-sion and real occlusion outperform other methods, which also demonstrates the efficiency and robustness for the proposed methods.(3) Furthermore, we consider the scenario that neither supervised information nor occlusion training samples are available. Considering the deficiency of independent assumption in standard sparse representation and local geometric structure simultaneously, we propose to learn robust sparse face representation via graph embedding structural sparsity. The standard sparse representation assumes that the coefficients are independent, which is not suitable for real application. Because the sparse representation of similar signals should be similar. Considering the geometric structure of faces manifold, we adopt locally group sparse graph to represent the relation between different subjects, and embed this graph into the sparse representation framework, then obtain an structural sparse model. The problem is decomposed into many sub-problems. A fast proximal algorithm is proposed to solve the sub-problem. The experimental results show that learned sparse representation is more robust, discriminant and achieves comparable results during the face clustering.