Face recognition is an important research direction with well theoretical research value and practical value in pattern recognition area. Due to it is not intrusive, easy to collect information and low cost, automatic recognition is widely used in many areas such as Customs inspection and video surveillance and so on. Sparse representation is an important breakthrough in the field of face recognition, which can effectively improve the robustness of face recognition.In this paper, the robust sparse representation based on classification method is used for face recognition. After studying the sparse representation based on classification methods, this thesis researches sparse representation based on classification without training dictionary and sparse representation based on classification training dictionary respectively; and then discusses the impact of the discriminant training dictionary on classification performance, and improves two kinds of sparse representation classification, respectively, which achieves the purpose of improving the recognition rate in face recognition applications. Researching work includes the two following aspects:1. Sparse representation and its face recognition applications. Firstly, study the basic ideas and methods of sparse representation; Then verify the robustness of the classification based on sparse representation for face recognition; Secondly, focus on improving recognition performance of the training dictionary for sparse representation by two dictionaries training methods of Fisher and Metaface, and experiment prove that the training dictionary can effectively improve the robustness of face recognition.2. Face recognition based on sparse representation methods. This thesis improves two recognition methods without training dictionaries and with training dictionaries, specific work as followings:1) In the algorithm of face recognition without training dictionaries, the paper directly uses training sample set to reconstruct test image sparsely, and then use the reconstruction error for recognition. In order to improve the discrimination of coefficients, a method which combines locality-constraint with the sparse constraint is used to improve the algorithm in sparse reconstruction. Ensuring sparsity of coefficients, the improved algorithm makes the coefficients tend to choose the training samples which have the same subspace as test image, thus enhancing the discrimination of coefficients. The experiment verifies that this algorithm improves the face recognition rates.2) In the face recognition algorithm with training dictionary, a method which is used to improve the training dictionary by using distance-constraint. Distance-constraint can fully exploit the local characteristics of the data, and filtrate atoms of sparse dictionary from the same class, and achieve a discriminated structured dictionary, which contributes to improving recognition performance. The dictionary is robust clustering of training samples, which can completely represent the training samples by fewer atoms. The experiment verifies the discrimination of the improved dictionary, which obviously improve the face recognition rate compared to other algorithms.3) Since the objective function of improved training dictionary is convergent, the closed form solution can be used to update dictionary and sparse coding, thus avoiding complex 1l-norm optimization problem, which effectively reduce computation of dictionary training. |