Image Local Structure Based Feature Extraction And Classification For Face Recognition

In the last few decades,face recognition is always an active research topic in the field of computer vision and pattern recognition. Thus it has attract significant attention and obtained tremendous achievements. However, most face recognition methods heavily rely on the number of training samples. Therefore, their performance drops significantly or they even fail to work when encountering single sample per person (SSPP) problem.In this paper, we develop some novel feature extraction techniques and classifiers by using image local structure to solve SSPP problem. The proposed approaches not only achieve good performance but also have great robustness to expression, illumination, occlusion and time variation. The main work and research achievements are as follows:(1) A local feature extraction approach called Weber Local Binary Pattern (WLBP) is proposed, which is based on the psychology law of Weber’s law. Specifically, WLBP consists of two components:differential excitation and LBP. Differential excitation extracts perception feature by Weber’s Law, while LBP can describe local features splendidly. By computing a two-dimensional histogram from the two components, we obtain a more discriminative statistical histogram feature. The perception feature extracted by differential excitation makes WLBP more robust to nosie and illumination.In addition, WLBP has already been successfuly applied in texture classification, face recognition and eye states detection.(2) Local feature based face recognition approaches for single sample per person problem are proposed. To solve face sketch recognition problem which is a special case of SSPP problem, we propose local geometry feature extraction apparoach. We first asumme that small image patches in the photo and sketch images form manifold with similar local geometry in two different image spaces. Then local geometry structure between local image patches can be used as feature to convert face photo and sketch into the same feature space that reduces the modality gap between photo and sketches. Furthermore, we also make use of local geometry feature, WLBP, LBP, WLD and Gabor to solve SSPP problem in face recognition. Compared with global feature, the above mentioned local features are more robust to SSPP problem.(3)Local similarity based subspace learnig algorithms are proposed. The performance of the traditional subspace learning methods such as PC A, LDA, LPP drops significantly or even fail to work when encounerting SSPP problem. To sovle this problem, we divide the face image into local blocks, and classify each local block, and then integrate all the classification results to make final decision. To classify each local block, we further divide each block into overlapped patches.and propose local similarity assumption according to the local structure relationship. Based on the local similarity between the overlapped patches in a block, we propose a seires of local similarity based subspace learning methods: local similarity based linear discriminant analysis (LS_LDA), local similarity based median discriminant analysis (LS_MDA), local similarity based principal component analysis (LS_PCA), local similarity based locality preserving projection (LS_LPP) and local similarity based marginal fisher analysis (LS_MFA). These local similarity based methods not only sovle the SSPP problem but also have great robustness to illumination, expression, occlusion and time variation. Moreover, motivated by the framework of lienar graph embedding algorithm, we further propose a unified framework called local similarity based linear lienar graph embedding (LS_LGE) to summarize LS_PCA, LS_LDA, LS_LPP and LS_MFA into the framwrok. The LS_LGE framework has good generalization ability and can generalize more local similarity based subspace learning algorithms.(4) Local structure based sparse representation classification (LS_SRC) and local structure based collaborative representation classification (LS_CRC) algorithms are proposed.The performance of sparse representation based classification (SRC) drops significantly when encountering SSPP problem. To solve this problem, we propose local structure based sparse representation classification (LS SRC). We divide each face into local blocks which consist of overlapped local patches. Then we assume that those overlapped patches in a local block lie in a linear subspace. The subpace assumption not only reflects the local structure relationship but also makes SRC feasible to SSPP problem. To lighten the computing burden of LS_SRC, we further propose local structure based collaborative representation classification (LS_CRC). Considering the fact that some irrelevant training samples will disturb collaborative representation and have adverse influence to classification, we propose local structure based two-phase collaborative representation classification (LS_TPCRC) and local structure based multi-phase collaborative representation (LS_MPCRC) which eliminate those useless training samples and by sample selection. The sample selection technique not only guaranttes the collaborative representation ability but also produces sparseness in a supervised way. Moreover, to further improve the performance of LS_SRC and LS_CRC, we also propse confusion matrix based Bayesian inference which makes use of confusion matrix to describe the error of classifier and derives more accurate probability of the test sample belonging to each class by Bayesian inference.