Construction Method Of Principal Component Networks And Its Application

People often obtain information through words rather than image and audio/video on the Internet before entering the era of information, because of the huge size of data in image and audio/video as well as the computational complexity when processing them. With the develop-ment of computer hardware and software technology, it is possible for us to process the huge volume of complex data such as image and audio/video. Images have received significant at-tention because they contained massive and diverse of data. Through image recognition, people have developed intelligent systems like driver assistance systems, security system and Intelli-gent Transportation System. Feature extraction, which will reflect results directly, is the most significant part in image recognition.Convolutional networks are one of the frontier areas of research in image recognition, it has made many breakthroughs in the domain of image classification. Convolutional networks stack multistage similar structures, where each stage includes three layers:filter bank layer, nonlinear layer, and pooling layer. Convolutional networks share most of their parameters so they can simplify the structure of the network and avoid overfitting in a certain degree. Recog-nizing image with features extracted by convolutional networks enjoys many advantages, like being robust to illumination variation, stable to rigid deformation or nonrigid deformation and invariant to scale.The thesis proposes several convolutional networks including Kernel Principal Compo-nent Analysis Network(KPCANet),Independent Component Analysis Network(ICANet) and Robust Principal Component Analysis Network(RPCANet), which are constructed by combin-ing several feature extraction algorithms and support vector machine. We have studied how these feature extraction algorithms affect the image classification tasks.KPCANet is a convolutional network using the kernel principal component analysis(KPCA) algorithm to extract features. Various forms of KPCANet are constructed by implying various kernel functions and they are verified with multiple kinds of images including face images, handwriting images, texture images and objects images. Besides, the impact of kernel function and also the parameters on KPCANet is tested with MNIST handwriting dataset.Traditional convolutional networks are unsatisfactory in extracting features of texture im-ages, while independent component analysis(ICA) is suitable to extract texture features. ICANet is constructed with ICA and it is tested with multiple texture image dataset.Robust principal component analysis(RPCA) could recover the severely corrupted im-ages,therefore, we have also constructed RPCANet with RPCA to extract features from severely corrupted images and test it with AR face dataset.Experiments show that KPCANet, ICANet, and RPCANet can extract image features ef-fectively. Besides, KPCANet is suitable to extract a various type of images and it is robust to illumination variation and stable to slight non-rigid deformation. ICANet is efficient in extract-ing textures features and robust to scale transform. RPCANet is efficient to extract features of severely corrupted images and it is robust to illumination variation as well.