A Study Of Image Classification Based On Transform Domain Feature And Deep Learning

With the rapid development of Internet, the image, as the basic form of visual information, has become a common carrier of information after text, and the mount of images are growing fast. Image classification is a technology for image understanding by extracting features of different images and recognition, which has very important significance for the development of society and the reality of work.In this dissertation, we study the visual feature extraction of image classification problems, and focus on the frontier method called deep learning. The traditional deep learning architecture has many parameters for training, and makes training very slowly. The dissertation proposes two new frameworks for image classification to face this problem. In contrast to the traditional deep learning architecture using raw pixel as input, we firstly apply prefixed cosine or wavelet filter bank to transform input to a new transform domain. The description of original images will have good invariance and distinguish properties. Then construct a deep network above this description to learn unknown far more complex sources of variability. The main work and innovations of this dissertation are as follow:1. Propose a novel framework by combining discrete cosine transform and deep networks for high speed object recognition system. The main idea of this approach is to apply the discrete cosine transform to reduce the information redundancy and select only a limited number of the low-frequency coefficients to feed into a deep network directly instead of raw pixels. And then the deep network is trained to learn good high-level representations of data in frequency domain in unsupervised fashion, and then for final image classification.2. In view of the traditional scattering description which is invariant to translation, rotation, scale and linear transformation of image, this dissertation proposes a novel framework by combining scattering transform and deep learning architecture. On one hand, wavelet scattering networks may provide the first two layers of these general deep architectures. The invariance and stability properties of scattering operators have the capacity to eliminates translation or rotation variability of input and simplify the learning task of subsequent layers, since they map image patches into a regular manifold. On the other hand, construct a deep network above these scattering coefficients can learn unknown far more complex sources of variability which the scattering transform cannot describe.These two frameworks use predefined filters to replace the filters at the bottom layers of the network. On the one hand, it can avoid learning the parameters of these filters from data. On the other hand, it simplifies the learning task of subsequent layers. In contrast to the traditional deep learning architecture using raw pixel as input, these two frameworks cost less training time and computing resources, and will provide a more effective solution for general users to apply deep learning technique. Therefore, the works of this dissertation have wide range of application scenarios.