Study On Key Technologies Of Bow Model Based Image Classification

Image Classification is a research focus in Computer Vision, while machine learn-ing, which is the theoretical foundation and methodology of Image Classification, is developed due to the complicated problem in real application. Although it is hard for computer to fully understand the semantics of images, but, in recent years, "Bag-of-Words" model has been successfully introduced into Image Classification. Further, a series of effective algorithms have been proposed for Image Classification so that we gained a great progress in this task. In this thesis, we focus on the study of key steps of "Bag-of-Words" model, that is codebook learning and feature pooling.Firstly, in order to enhance traditional clustering approaches adopted in codebook learning, we present an incremental neural network learning algorithm. It can learn appropriate visual words set via online approach and provide richer information for feature encoding since the codebook is a network. Furthermore, we design subgraph-based coding method, which effectively encodes local features by the relationships among visual words, leading to better performance. Experimental results show that efficiency and accuracy of Image Classification are improved by combining these two methods.Secondly, after analyzing drawbacks of existing sparse representation based code-book learning algorithms, we introduce self-paced learning mechanism into codebook learning, and propose a framework for sparse representation based codebook learning, which uses the easy samples to train the codebook first, and then iteratively introduces more complex samples in the remaining training procedure. Experimental results vali-date the method and show that it can improve classification accuracy.Finally, we check second-order feature pooling carefully, and then we present a new image second-order representation combining low-dimensional representation and Graph-embedding discriminative analysis on Riemannian manifold to overcome the problem of high-dimensional image representation from second-order feature pooling. Experimental results show that our method can effectively reduce the dimension of image representation and keep high classification accuracy.