In this paper, we encounter the problem that the training image is neutral sample obtained from a normal condition rich in light, and there are kinds of variations containing illumination, expression, poses and occlusions in testing samples. Sparse representation based face recognition is analyzed and applied to face recognition in our method, and a deeply reseach on complex face recognition is proposed including following sections:1. To improve the representatation ability, a class-specific dictionary learning method is propoed learning sub-dictionaries for each class such that every atom own its label. In the learning process, by applying a constraint that leads to the structure dictionaries to be incoherent, the learned sub-dictionary is able to reconstruct the data from the same class but can not well represent images owing to different classes. As a result, when a testing sample comes in, the sparse coefficient as well as the reconstruction error are untilized for verifying the final label, which improves the accuracy.2. Aming at the complex variations existing in query samples, a shared dictionary learning algorithm is presented to get the shared features across different classes. Through proposed shared dictionary learning method, the variations between the same classes are catched. The combine of shared dictionary and proposed class-specific dictionary can effectively represent the testing samples that contain many variations.The effect of feature for dictionary learning and classification is analyzed, and the experimental results verify that using more robust Gabor feature leads to a better performance. Furthermore, by comparing the classification performance between the proposed methods and other classical face recognition algorithms, the effectiveness and superior of the proposed methods are demonstrated by the experimental results.3. To overcome the limitations existing in face recognition that the number of training samples is not sufficient enough and the testing samples have variations(illumination,expression,occlusion,pose), a new face representation algorithm by two dictionaries including class-specicfic and intra-class variation dictionary is proposed. To avoid the influence of the manually designed dictionaries which are somewhat person-specific, we present a new intra-class sparse variation dictionary. The combination of class-specic and intraclass variation dictionary enables our method to achieve good performance on the AR, CMUPIE, FERET and Extended Yale B databases. |