Complex Image Classification By Sparse Feature Learning

From now on, the Internet Age has come, which brings great changes to us in learning, creating and transmitting the knowledge. Recently, people are always mining useful information by the text retrieval in the internet. However, with the development of internet techology, there are more and more image and video in the internet and scholars are paying their attention to the content-based image and video retrieval. As image classification is one of the most important topics in image analysis and understanding, it has become a hot research topic in this. This paper focus on solving this problem.In recent decades, sparse representation has become a very hot topic in image processing, image analysis and understanding. Sparse-representation-based methods have been proposed in almost every fields in image processing, image analysis and image understanding. More and more research results show that sparse representation is related with images. Therefore, all the proposed classification methods in the paper are based on sparse representation. Now, sparse-representation-based image classification methods faced up three main problems:(I) how to design a better method to learn a dictionary for sparse classification;(II) how to design a better optimization algorithm to solve the sparse models efficiently;(III) how to design a better algorithm to classify the learned sparse representation features. Based on this issue, this paper is written in the three aspects.In Chapter 2, a new algorithm---Incomplete variables Truncated Conjugate Gradient method(ITCG) is proposed to solve the sparse representation models with 1- norm. By adjusting the parameters, two specific algorithms are presented, i.e. ITCG-vs for very sparse reconstruction and ITCG-nvs for not very sparse reconstruction. To make full use of the sparse nature, ITCG can obtain them efficiently. The experiments show that the two algorithms of ITCG(especially ITCG-nvs) are much faster than competing methods in sparse reconstruction. In addition, it has been shown that ITCG-vs can converge after finite iterations under some decent conditions.In Chapter 3, a new sparse coding method---Spectral-Spatial Sparse Coding(S3C) is proposed for Hyperspectral Image(HI) classification. To improve the performanc, a spectral-spatial dictionary is learned by a two-stage affinity propagation clustering method based on spatial neighbors, and then a local sparse coding method is used to structure spectral and spatial sparse features(S3C features). As S3 C features are sparse and include rich information from both spectral and spatial domains, it can be classified efficiently and accurately by linear support vector machine.In Chapter 4 and 5, two sparse-representation-based algorithms are proposed to classify the face images. They are named as Evidential Reasoning based classification algorithm(ERC) and Sparse-Representation-based Classification algorithm with Outlier-clearing Strategy(SRCOS). Because they are designed to classify the images with noise, their common characteristics are to remove as much of adverse effects from noise in the training samples as possible. The difference between them is that ERC employs evidential reasoning theory and builds belief rule base to decrease the class noise in the training set while SRCOS learns a weight for every training image to identify if the training image is occluded. With these strategies, ERC and SRCOS can obtain the better classification accuracy under the conditions of class noise and occlusion noise, respectively.