Font Size: a A A

Research On Face Recognition Under Incompletely Sampling Based On Data-oriented Sparse Representation

Posted on:2017-02-08Degree:DoctorType:Dissertation
Country:ChinaCandidate:S H ZhaoFull Text:PDF
GTID:1108330503482654Subject:Electronic Science and Technology
Abstract/Summary:PDF Full Text Request
Face recognition has been widely researched and applied in real life for its non-contact, easy collection. The main applications of face recognition contain check in, access control system, surveillance, public security and etc. Most of existing methods have achieved satisfying performance in controled conditions, but many challenges in practice would drop the performance, such as: occlusion caused by illumination changes, scarf and sunglasses, small samples, various poses and expressions. All these problems will cause the mission of face information, and will drop the performance of existing methods. In this paper, we call the face recognition in these conditions as incompletely sampling face recognition. Sparse representation based classification(SRC) has achieved great success in image recognition, so we try to improve the performance of incompletely sampling face recognition based on SRC. The main contributions of this paper are summarized as follows:(1) Weighted neighbor classes based block sparse representation and discriminative decomposition based block sparse representation for occluded face recognition are proposed. Weighted neighbor classes can select samples from classes with small distance with test sample, thus the computation cost is reduced and the recognition rates are improved. Besides, we try to do decomposition on the training samples to separate the occlusion part and the training samples are decomposed into common part, condition part with low rank and sparse error. Principal component analysis(PCA) is performed on common part and condition part respectively to obtain the projective matrix. The final identification is conducted on the projective subspace by BSR(Block Sparse Representation).(2) To eliminate the effect of occlusion, we divide image into some modules and propose different strategy to compute the modular weight. For the modules with occlusion, we assign small weights and for the clean modules without occlusion we assign high weights. Thus, the effect of occlusion is declined. Firstly we try to weight modules by Fisher rate and the final classification is performed by SRC, experiments on some face databases verify that this method can improve the recognition rate to some extent. Besides, we try to use the sparse residual to weight the modules and estimate the occlusion part. Finally, we joint the Fisher rate and sparse residual to assign the modular weights, and the experiments results show great improvements in recognition performance.(3) To detect the occlusion more precisely and realize the occlusion face recognition when there is no occlusion in training set, we propose two occlusion detection methods on the pixel level: pixel-level occlusion detection based on sparse representation and block sparse recursive residuals analysis for face recognition. In the first method, we estimate each pixel occluded or not based on the residual. Then, we compute the residual of each class and make occlusion estimation based on each classical residual, the final occlusion is estimated by each class’ s estimation. Thus the occlusion pixels can be detected precisely. In the second method, images are divided into two modules and the module with higher sparsity is used to reconstructed the global image and the residual is used to estimate the occlusion pixels, thus we can weight each pixel and identify the test image. The occlusion detction methods on pixel-level can eliminate the effect when the occluded block contains unocclusion pixels, thus the recognition performance can be impoved.(4) We use the kernel trick to extend the BSRC(Block Sparse Representation based Classification) and propose kernel block sparse representation based classification(KBSRC). KBSRC projects the samples into the kernel sunspace, so the nonlinearli-separable original subspace can become linearly-separable. Besides, in the kernel subspace we take the structural information into the recognition by BSRC to improve the recognition rates.
Keywords/Search Tags:face recognition, incompletely sampling, data-oriented, sparse representation, occlusion detection, kernel trick, discriminative analysis
PDF Full Text Request
Related items