Research On Face Recognition Via Robust Regression

Face recognition is a hot research topic in recent two decades. However, this problem is very complex, since the facial images have large inner-class distance and small outer-class distance under variations. Recently, linear representation theory based methods obtained promising results, so to solve the face recognition problem by applying the theory is one of the most popular research paradigms in the area. These methods assume that a facial image can be represented linearly via the images from the same class. Nevertheless, except identity, facial image also contains some interferences such as emotion, gesture, illumination, occlusion and so on, and these interferences may change the location of an image in feature space. Therefore, the image may not be represented linearly by the training images from the same class.To solve the problem, this thesis researches linear representation based methods for robust face recognition. Firstly, a robust regression based method is proposed which employs Huber loss to suppress noises in facial images. In addition, this thesis studies the relation between noises and representation residuals, meanwhile an error detection approach based on multi-step linear representation is proposed. Furthermore, an adaptive error detection method is proposed which can make an appropriate adjustment according to the noise in a face image. The highlights and main contributions of this paper include:1. A robust supervised sparse representation based method is proposed. The method employs Huber loss function as fidelity term, so the loss function can alleviate the influence of noises in facial image. Therefore, the robustness of the linear representation method is boosted. In addition, the supervised sparse representation is used to control the sparsity of coding, so the complexity of solving the model is greatly reduced.2. This thesis analyzes the distribution of the noise in representation error, and a multi-step linear representation based on error detection approach is proposed. The approach dose not have to assign a weight for a common error (i.e. the error value is not too big nor too small), so the reliability of error detection is improved.3. This thesis analyzes the relation between the noise proportion in facial image and the sparsity of coding vector and proposes an adaptive error detection approach. The method dose not have to presuppose the ratio and distribution of the noises in face image, and it can adjust according to the types of the noise automatically. According to our experimental evaluations, this method improves the generalization and robustness of linear representation model.