Research On Highly Robust Face Recognition Against Occlusion

Face recognition technology has found wide applications due to the property of non-mandatory,non-contact,concealment,concurrency and high efficiency.However,in the practical application of face recognition,there are some difficult problems remaining unsolved and the occlusion problem is a prominent one.The various occlusions make the intra-class variations larger than the inter-class variations in the sample set,which blurs the category attribute of the sample and makes classification more difficult.To address this problem,this work thoroughly studied the occlusion problem and proposed three robust face recognition methods with local occlusion.The main contributions of this work are summarized as follows:1.The theory of sparse representation and collaborative representation are studied in detail,including the basic concepts,models and algorithms.Then the applications of sparse representation and collaborative representation in face recognition are described respectively.2.A method of occlusion face recognition based on block image similarity weighting is proposed.The similarity between blocks is measured by collaborative representation coefficient and based on which a weighting matrix is constructed.Then the Weber features extracted from the sample set are mapped to kernel space where the test samples are classified with the weighted collaborative representation classification.3.An occlusion face recognition method based on residual weighting is proposed.Based on the sparse representation of image blocks,the weight vector is constructed with the sparse coefficients and the reconstruction residuals.Then the Weber local descriptor extracted from the sample set is mapped to the kernel space where the test samples are classified using the residual weighted collaborative representation classification.4.A face recognition method based on occlusion area removal is proposed.The method takes into account the fact that the occlusion objects are generally different from the human skin in terms of intensity.Therefore occlusion area can be detected from the difference image of the training image and the test image.After removal,sparse representation classification is employed to classify the non-occlusion area.Experiments on the popular AR face database and the Extended Yale B face database show that the above proposed methods exhibit strong robustness and high recognition rates for face images with local occlusions.