Based On Subspace Learning Complex Scene Multi-pose Face Recognition

Face recognition is a hot issue in the biometric identification. There are broad applications, such as the public safety, the financial safety, the augmented reality, and the man-machine interaction etc. In recent years, the technology of face recognition got a large progress. However, because the face images are affected by several factors, it is difficult to design a general algorithm to solve the face recognition problems under all circumstances. The face images have a lot of redundant data resulting from its high dimensions and non-structural property. Owing to the complicated structure of the face images, the recognition rate will be bad if the original image values are directly used to classification. The subspace methods have been the most popular methods of face recognition, which build subspace vectors based on the samples, and project the face image into the subspace eigenvectors to extract its feature. This procedure not only lessens noise but also improves the speed of recognition. In simple scenes, the methods of subspace can achieve good recognition performance. However, the performances of the existing subspace algorithms drop rapidly in the complex reality environment. This paper researches mainly the face detection and recognition algorithm based on the subspace-based methods. The aim is to improve the performance under the complex scenes. The main contributions of the dissertation can be noted as following:(1) The research of face detection based on LPP-AdaboostTraditional Adaboost algorithm has a high real-time performance. However, the performance would be sharply down in the complex background. This paper proposes a human face detection algorithm based on LPP-Adaboost. Firstly LPP (Local Preserve Project) is used to extract the features of face images under various conditions, such as the poses, the illuminations, and the expressions. LPP is the linearization of the nonlinear Laplacian Eigenmaps, which has both advantages of the linear methods and the nonlinear methods. Compared with the linear method, LPP is better to describe the manifold structure of human faces; and compared with the nonlinear method, LPP can get a projection subspace which is used to classify new samples. The algorithm extracts LPP eigenvector set from the training samples, and chooses the well classification eigenvectors as the weak classifiers from the whole LPP eigenvector set. Then the weak classifiers are boosted to the strong classifiers by Adaboost according to the classification accuracy. Finally, the ordered strong classifiers are cascaded to classify the face image patches. Experiments results showed that the LPP-Adaboost algorithm has better performance than other linear methods in complex background, and has fast detection speed.(2) The research of face recognition based on subspace methodThe subspace method is an effective way to overcome the curse of dimensionality, whose main idea is to project the samples which distribute loosely in the high dimensional space into the low dimensional space by the linear or nonlinear methods. Such projection makes the samples more compact, more favorable to be classified, and has less computational complexity. This paper proposes a face recognition algorithm based on LBP-SR algorithm, which uses the scale invariant LBP operator to extract multi-scale features. Firstly, Gaussian filtering and down-sampling are used to build the image pyramid, from which LBP operator is adopted to extract the LBP features of each sub-image. Then, the multi-scale LBP features are feed as the input of the spectral regression to extract the eigenvectors in the projection subspace. Finally, KNN is used to classify the label of sample. The multi-scale LBP features are rotation invariance and translation invariance, which replace the image pixels to get the advantages of the computational complexity. In addition, in the spectral regression feature subspace, the affinity matrix can describe the relations of the samples. And the regression of constructing projection space need not to solve the density matrix explicitly which is the problem of the traditional subspace methods. Experiments results indicated that our algorithm has better performance in the complex background with fast recognition speed.(3) The implementation of the face recognition based video surveillance systemBased on the study of algorithms above, we implement the video surveillance system based on the faces detection and the face recognition. In the system, the Hikvision Video Surveillance Camera is used. The system has three modules:the detection and extraction of video object module, the face detection module and the face recognition module. The Adaboost based method and the LPP-Adaboost based method are used to implement the face detection module. PCA, LDA, LPP, LBP, LBP-SR methods are applied to implement the face recognition module. The system 'utilizes the components model to well define the general training classifiers, the recognizer and the performance measurement that can be conveniently embedded, measured and compared with other methods. A unified data access interfaces is designed to adapt various popular standard face datasets. So the video surveillance system provides an open source platform for the face recognition, which not only can be commercialized, but also can be used for the algorithm researches.