Research On Occlusion Face Recognition Based On Low Rank Matric Recovery And Sparse Representation Classification

With the huge breakthroughs in face recognition theory and the improvement of computing capabilities,face recognition technology has received significant attention and been used in various aspects of life widely in recent years.Nevertheless,most algorithms existed can only achieve good recognition results under strict laboratory conditions.However,the recognition accuracy of these algorithms are seriously reduced in real application scenarios due to the effects of illumination pose,expression,and occlusion on the face image.Therefore,based on the theoretical basis of low-rank matrix recovery and sparse representation,this thesis focuses on the problems of poor real-time performance and low recognition accuracy of occlusion in face recognition algorithms.The specific researchs are as follows:1.For the problem of inaccurate of low-rank matrix recovery and small samples when the sample data comes from different subspaces,a collaborative representation classification occlusion face recognition algorithm based on discriminative low-rank matrix recovery is proposed.Firstly,a structurally unrelated regularization constraint term is introduced into the low-rank matrix recovery to recover low-rank training samples from the contaminated training samples effectively.Then,with a low-rank projection matrix is learnt between the original defaced training samples and the low-rank training samples,the correction is performed by projecting the defaced test samples to the corresponding low-dimensional subspace.Finally,the recognition result is obtained by the collaborative representation and classification method to complete the classification of the modified test samples.Experimental results show that this method not only alleviates the problem caused by small samples,but also improves the effectiveness of occlusion face recognition.2.For the problem that the overall features cannot effectively represent occlusion when both the training sample and the test sample are occluded,an adaptive sparse representation occlusion face recognition algorithm based on Gabor features is proposed.Firstly,this method is based on the algorithm proposed in Research Point 1.It restores clean low-rank training sample images and enhances the discrimination of the recovered samples.Then,the Gabor transform is performed on the recovered low-rank face image,the original training samples,and the test samples to obtain the corresponding Gaborfeature vector and constructing a more compact feature dictionary.After that,the principal component analysis is used to obtain the transformation projection matrix of the low-rank linear subspace in which the low-rank training sample Gabor feature dictionary is located.Moreover,the Gabor feature dictionary corresponding to the original training samples and the test samples is projected to the same linear subspace through the transformation projection matrix and performed the adaptive sparse coding.Finally,the classification recognition is realized with adaptive sparse representation classification.The experimental results show that this method has a higher recognition accuracy and stronger robustness for occlusion face recognition,and is more suitable for practical applications.