Research On Quaternion Autoencoders And Its Applications

Along with the further exploration of the internet world, the kinds of signals in contact are becoming varied. For example, we live though the beginning of text message, to subsequent development of acoustic signal, and then the valuable image information. The approach of human being getting information become more and more convenient, and the obtained content become more and more meaningful. The significant image information has turn out to be the main way in information transmission on internet due to its abundant representation. But we will encounter some problems when we face large scale image data. For example, human beings cannot recognize large scale of image accurately and speedy due to the difference between ability of each people and the redundancy of image information. So it is urgently necessary for the recognition of image by using computer. The conventional way of image recognition will clean the redundancy information in image information through feature extraction algorithm, and then recognize image via these extracted feature. The traditional image recognition algorithm use principal component analysis (PCA), linear discriminant analysis (LDA), quaternion principal component analysis (QPCA), etc. to extract features from identifying image. Since these type of feature extraction algorithm are artificially specify to fitting the different types of image, the efficiency of algorithms are very low. So we cannot use these algorithm extract features automatically.A branch of deep learning, whose name is autoencoders, developed rapidly in recent years. It does not specify how to extract features from image like the above algorithms, instead of using a self-learning approach to learn exact feature from a given image data. And then using these extracted features in image recognition field. The Principal Component Analysis Network (PCANet) is a new type of autoencoders. In this thesis, we will introduce the architecture and analysis its theory in detail. The theory of PCANet can be divided into three parts: convolutional neural network, filter bank, and histogram. The filter bank in PCANet is the main research area in this thesis. Following with the color image prosperous in life, we also do complete experiments to test the performance of PCANet on various color space.It is worthy to note that quaternion representation of color image provoke widespread interest of many scholars. In this thesis, we will introduce the quaternion algebra, and then analyze the color image application of QPCA. At last, QPCA are used as a filter bank to replace that of PCANet. The new type of network are named as quaternion principal component analysis network (QPCANet).We also proposed two new algorithms, which are Multilinear Principal Component Analysis Network (MPCANet) and Multilinear Discriminant Analysis Network (MLDANet) respectively, to extract appropriately feature from tensor object. The performance of MPCANet and MLDANet are verify on UCF11 video database and has demonstrated that they are better than conventional tensor algorithms like Multilinear Principal Component Analysis (MPCA) and Multilinear Discriminant Analysis (MLDA).