Research And Implementation Of Lp-norm Based Generalized Principal Component Analysis Methods Of Face Recognition

Face recognition based on a person’s facial features attracts extensive research interests in the Artificial Intelligence field. Feature extraction is the key technique of face recognition, which can remove redundant information from the face image and extract the main features. The principal component analysis(PCA) is the most classic feature extraction method, however PCA and derivative is easily affected by facial expression, illumination variation, occlusion and other factors. In this thesis, PCA, two dimensional principal component analysis(2DPCA) and bidirectional two-dimensional principal component((2D)2PCA) is in-depth analysised., Several methods are proposed to improve the recognition performance. The main research contents and innovative work in this paper are as follows:(1) Feature Extraction Algorithm of PCA with Lp-norm Constraints(Lp-RM-PCA)For the methods based on PCA were easily affected by sample mean, a new feature extraction algorithm with Lp-norm constraints(Lp-RM-PCA) is proposed. The model is able to avoid the effect of the sample mean value on model’s optimization. An iterative algorithm is used to solve the model and to obtain the local optimal solution. The experiments show that the recognition performance of the proposed method is effective.(2) 2-Directional 2-Dimensional PCA with Lp-norm Constraints for Feature Extraction2-Directional 2-Dimensional PCA with Lp-norm Constraints for Feature Extraction((2D)2PCA-Lp)Two-dimensional principal component analysis(2DPCA) is sensitive to outliers and cannot obtain sparse projection vectors. In this paper, a new optimization criterion based on the L1 norm objection function and Lp-norm constraints((2D)2PCA-Lp) is proposed. An iterative algorithm is used to solve the proposed model and the local optimal solution can be achieved. It is found that one can obtain the sparse projection vectors if the parameter p in the model approaches 1. Experimental results on Yale, NYU_UMIST and ORL face databases demonstrate that the proposed method is much more effective than previous methods, especially in handling the contaminated data.(3) Sample Weighted Method Based on Lp Norm Constraints of(2D)2PCA for Face Recognition(Lp-RM(2D)2PCA)Considering the fact that 2-directional 2-dimensional principal component analysis is easily affected by sample mean, a new optimization model(Lp-RM(2D)2PCA) is proposed and an iterative algorithm is used to solve the model. This model can not only effectively remove the influence of sample mean, but also weight the pairs of samples using Gaussian functions. By choosing a proper parameter p in the proposed model, the sparse optimal projection vectors can be obtained, which in some degree alleviate the problem of the partial occlusions in face images. The experimental results on NYU_UMIST, Yale and ORL face databases show that the performance of the proposed method can be improved by 1%-4% compared with previous methods.(4) Design and realization of trajectory data generalization systemBased on the study in the thesis, a face recognition generalization system is designed and realized. The proposed methods are embedded into the system. The system also provides a good visualization and operability, which better presents the recognition results.In short, the experimental results show that these proposed methods(Lp-RM-PCA,(2D)2PCA-Lp, Lp-RM(2D)2PCA) are less sensitive to outliers and have better performance than some other models with outliers in face image.