Research On Face Recognition Algorithm Based On Manifold Learning And Collaborative Representation

Face recognition has a large number of applications, including security, person verification, internet communication, and computer entertainment. Face recognition, as one of the primary biometric technologies, becomes more and more important. It has great application value in many fields such as public safety and national security. Compared with other biometric identification methods, face recognition has the advantages of being natural, friendly and convenient, and has gained attentions by more and more researchers, who get a lot of research achievement. However, because face structure is varied, face is affected by illumination and face may be occluded, face recognition is very challenging and worthy of further research.Feature extraction and feature dimensionality reduction are two key steps in face recognition. How to effectively extract face feature is one of the key issues in face recognition. At present, when face is occluded, the result of face recognition is not ideal. In order to develop the robust and real-time face recognition system, a lot of problems need to be solved. Now face recognition usually requires sufficient training samples per person to obtain better results, but in many specific scenarios, there is only a single image per person can be used for training, the result of face recognition with a single training sample per person is not satisfactory. How to solve the problem of face recognition with a single training sample per person becomes a practical problem. In this dissertation, these practical problems in face recognition have been studied deeply, and some new face recognition algorithms are proposed. In this dissertation, the main contents and contributions are described below:1. This dissertation proposes a face recognition algorithm based on Riemannian manifold learning. Isometric mapping(ISOMAP) can only be applied to intrinsically flat manifolds. Locally linear embedding(LLE) does not reflect the global structure of manifold. A manifold with a metric tensor defined at each point is called a Riemannian manifold. The basic idea of Riemannian manifold learning(RML) is that the Riemannian normal coordinates in the low-dimensional space represent points in the high-dimensional manifold to realize the purpose of the dimensionality reduction. Experimental results show that the recognition rate of RML is higher than that of ISOMAP and LLE.2. This dissertation proposes a face recognition algorithm using fusion of local binary patterns and locality preserving projections. Locality preserving projections(LPP) is a dimensionality reduction algorithm of the linear manifold learning and a linear approximation of the nonlinear laplacian eigenmap(LE). LPP constructs the adjacency graph to model the manifold structure of face space, and then finds a set of basis images. LPP is obtained by finding the optimal linear approximations to the eigenfunctions of the laplace beltrami operator on the face manifold. LPP reflects the intrinsic manifold structure of face space and has the clear projection matrix. Therefore, LPP is very suitable for face recognition, but LPP is simply used for face recognition, recognition rate is not high. Local binary pattern(LBP) is defined as a gray-scale invariant texture measure. It can efficiently extract the face texture feature which represents the local structure of face images, so LBP is very suitable to be used for the face feature description and recognition. This dissertation combines LBP with LPP and proposes a face recognition algorithm using fusion of LBP and LPP, named LBP-LPP. In the proposed algorithm, firstly, the uniform LBP operator is used to extract the face feature, secondly, LPP is used to reduce the dimensionality of the face feature vectors, finally, the nearest neighbor classification approach is used for classification and recognition. LBP-LPP can achieve very high recognition rates. Experimental results show that the performance of LBP-LPP is better than that of most of the other algorithms.3. This dissertation researches fusion of Log-Gabor wavelets and discriminant locality preserving projections for face recognition. For face recognition, feature extraction is very important. Gabor wavelets and Log-Gabor wavelets are both effective face feature extraction methods, but Gabor wavelets have shortcomings, Log-Gabor wavelets just compensate for the shortcomings, so Log-Gabor wavelets are more suitable to extract face feature than Gabor wavelets. This dissertation combines Log-Gabor wavelets with DLPP and proposes a face recognition algorithm using fusion of Log-Gabor wavelets and DLPP, named LGDLPP. In the proposed algorithm, firstly, Log-Gabor wavelets are used to extract the face feature, secondly, DLPP is used to reduce the dimensionality of the face feature vectors, finally, cosine similar classifier is used for classification and recognition. LGDLPP can achieve very high recognition rates. Experimental results show that the performance of LGDLPP is better than that of the other algorithms.4. This dissertation researches robust collaborative representation for face recognition. Sparse representation based classification(SRC) and the robust sparse coding(RSC) have been successfully used for robust face recognition, and show strong robustness to face recognition with occlusion, but the computational complexity of them is very high. Research finds that it is the collaborative representation but not the 1l norm sparse constraint that makes RSC suitable for face recognition. This dissertation proposes a face recognition algorithm based on robust collaborative representation(RCR). Similar to RSC, an iteratively reweighted collaborative representation algorithm is used to solve the RCR model. Compared with RSC, RCR achieves similar recognition rates but with significantly lower computational complexity. RCR and the other algorithms are compared in AR and Extended Yale B face database. Experimental results show that the performance of RCR is better than that of the other algorithms.5. Face recognition usually requires sufficient training samples per person to obtain better results, but in some specific scenarios, there is only a single image per person can be used for training. For face recognition with a single training sample per person, collaborative representation based classification(CRC) has significantly less complexity than extended sparse representation based classification(ESRC), but CRC gets lower recognition rates than ESRC. In order to combine the advantages of CRC and ESRC, this dissertation proposes extended CRC(ECRC) for face recognition with a single training sample per person. ECRC assumes that the intraclass variations of one subject can be approximated by a sparse linear combination of those of other subjects, and uses an auxiliary intraclass variant dictionary to represent the variant between the training and testing images. Experimental results show that ECRC outperforms CRC and ESRC in terms of both high recognition rates and low computation complexity. On the basis of ECRC, this dissertation proposes fusion of Gabor wavelets and ECRC(GECRC) for face recognition with a single training sample per person. In the proposed algorithm, Gabor wavelets are used to extract face feature, and then ECRC is used for classification and recognition. Experimental results show that GECRC is better than that of the other algorithms.