Research On 2D-LDA And PCANet Face Recognition Optimization Methods Based On Random Sampling Technology

With the vigorous development of pattern recognition,image processing and computer vision,the research achievements of face recognition technology have been emerging and integrated into people's lives.However,due to the diversity of facial expression changes,the instability of face images to environmental impact,limited computer resources and space and so on,the application of face recognition technology in complex environment is facing many challenges.Feature extraction is an important step in face recognition technology,and has been a research hotspot of scholars at home and abroad.In this paper,the feature extraction algorithm in face recognition technology is taken as the main research object.The 2D-LDA method and PCANet method in the face recognition feature extraction method are studied and further improved,so that the extracted features can be more robust to illumination changes and the computational complexity is not too high.Besides,the performance of the face recognition algorithm can be also improved.The main research work and results of the thesis are as follows:Firstly,to solve the problem that 2D-LDA subspace feature extraction algorithm is not robust to environmental changes such as illuminations,this paper analyzes the advantages of sub-image method,fuzzy set theory and random sampling technology in feature extraction method,and proposed an improved method named fuzzy 2D-LDA face recognition method based on sub-image and random sampling(RS_subimage-F2DLDA).The effectiveness of the improved method is verified by experiments.Secondly,by analyzing the filtering kernels of the principal component analysis network(PCANet),which is based on the deep learning method,using only several eigenvectors corresponding to the largest several eigenvalues,it is found that some discriminant information exists in the discarded eigenvectors.In order to make the best use of discriminant information in all eigenvectors,a random sampling based PCANet method(RS_PCANet)is proposed.The improved method improves the performance of face recognition algorithm by enhancing the diversity of filtering kernels.Thirdly,considering the dense characteristics of Densely Convolutional Network(DenseNet),the original image of the first convolution layer is added to the input of the second convolution layer in RS_PCANet to form a cascade of the input in second convolution layer,and a dense PCANet ideal(Dense_PCANet)is proposed.In addition,in order to preserve the geometric structure between pixel points of the image in the low-dimensional space,this paper added the smoothness constraint to the PCA mapping of PCANet method,and the ideal of smooth PCANet(Smooth_PCANet)is obtained.Preliminary experimental results indicate that the two ideas are feasible.