Application Of Block SRC Algorithm In Occluded Face Recognition

Since automatic face recognition technology has the characteristics of natural, friendly and directly, it is widely used in public safety information security, finacec and other fields in the1960s. However, as face images are non-rigid, they are easily affected by various internal or external factors, such as age, expression, racial, gender, disguise, occlusion, illumination, erosion and position. Thus, in face recognition problem, the key problem is designing an algorithm, which has the performance that can effectively weaken those interference factors. Sparse Representation Classification (SRC) is a recent proposed new algorithm, which different from traditional algorithms and has perfect recognition rates under occlusion, erosion, illumination and so on. Besides, SRC algorithm has nothing to do with feature extraction algorithm when the image space dimension achieves to a certain number. In this paper, Firstly, we detailly discuss the advantages and disadvantages of SRC and its improved voting SRC algorithm which is base on the whole sub-blocks; then, a complementary algorithm which named local SRC algorithm of voting SRC is presented. Although local SRC algorithm can make up for the shortage of mistaking recognition when the subjects of all sub-blocks are different from each others, if it is used alone, it will produce the sitiation of taking a part for the whole. Therefore, we consider the fusion algorithm that fusing voting SRC and local SRC algorithms together. From the review of recognition rates, fusion algorithm is best, comparing with voting SRC or local SRC algorithm in both outocclusion and occlusion face images recognition problems. In total, we work as follows:(?)The processing of image block. We mainly study two image block methods named uniform block and block base on characteristics. In uniform block, face image is viewed as a plane matrix no matter how the characteristic of image looks like, such as high nose and big eyes. Before recognition, firstly, a training image is uniformly divided into serval sub-blocks and each sub-block is storaged as column vector. Meanwhile, all sub-blocks are saved in a matrix. Then, all of the training images are divided in the same way and saved in other matrix. Finally, the testing image is partitioned in the same way. While in characteristic block, according to the principle of three court and five eyes, face images are partitioned into five sub-blocks, which are forehead, left eye and left eyebrows, right eye and right eyebrows, nose, mouth and jaw. Since the pixels of each sub-block are different, the five sub-blocks are storaged with five matrixes independently.(?) The structure of dictionary. The commonly structure of dictionary has two ways. One way is online learning, that is, in the process of face recognition, according to testing images, dictionary is adaptived choice. The other way is constructored by static. That is, before recognizing, dictionary is confirmed by us. In this paper, we will study Gabor+PCA dictionary which belongs to online learning and PCA dictionary or down-samlping dictionary, which belongs to static constructor.(?) The optimization solutions of l1-minimization. There are many sparse optimization algorithms such as classic Matching Pursuit (MP), Basis Pursuit (BP) and fast Truncated Newton Interior-point Method (TNIPM), Iterative Shrinkage-Thresholding (1ST) and so on. Since classic algorithms are slower than fast algorithms in running speed, in this paper, two fast dual Augmented Lagrange Multiplier (DALM) and Homotopy algorithms are considered.(?) The selection of distinguishing method. In this paper, we discuss three distinct methods based on sub-blocks of face images, which are called voting SRC, local SRC, and fusion of voting SRC and local SRC.In all, in this paper, the main problem is applications of block SRC in face recognition under occlusion. It contains four parts, which are how to block, how to constructor dictionary, how to select the optimization algorithms and how to confirm distinguishing methods. We select dictionary and sparse optimization algorithm are based on ORL and Yale face database. Whlie, AR face database is used to detect the performance of each block SRC methods. The experimental results show that, in uniform block, the way of partition can greatly affect the recognition performance. Meanwhile, no matter how to block, the fusion algorithm of voting SRC and local SRC algorithm is the best algorithm than voting SRC or local SRC algorithm.