Dictionary Learning Algorithms With Applications To Hyperspectral Image Classification

The past several years have witnessed the rapid development of the theory and algorithms of sparse representation and it has been successfully applied to a variety of problems in computer vision and image analysis. The classification performances of the classifier based on sparse representation depend on the choice of the dictionary, so how to learn a more representative dictionary from the training samples becomes the crux of designing a classifier induced from sparse representation.The dominant dictionary learning methods incorporate discriminative strategies into the objective function to enhance the discrimination of the dictionary. The discriminative strategies include linear predictive classification error, Fisher discrimination criterion, the label consistent and so on. These dictionary learning methods incorporate one discriminative strategy into the objective function, so the discrimination of the dictionary is enhanced in a sense. But if we incorporate different discriminative strategies into the objective function, the discrimination of the dictionary is enhanced in many ways and we can obtain a more discriminative dictionary and sparse codes. Hence, we do research into dictionary learning and emphasize particularly on incorporating different discriminative strategies into the objective function to enhance the discrimination of the dictionary and sparse codes. Experimental results show the effectiveness of the proposed dictionary learning models. Meanwhile, we apply the proposed dictionary learning methods to hyperspectral image classification (fusing the texture features with spectral features of the hyperspectral images and applying dictionary learning to hyperspectral image classification) and the experimental results further show the effectiveness of the proposed models. We have done the main tasks as follows:1. Propose a dictionary learning algorithm based on linear regression and Fisher discrimination, which is called LRFDDL. We incorporate the linear regression term and the Fisher discrimination into the objective function to get a more discriminative dictionary and one classifier. The linear regression term makes the predictive label and the actual label as close as possible, which is good for classification. Because the discrimination of the dictionary can be represented by the sparse codes and the Fisher discrimination criterion is imposed on the sparse codes so that they have small within-class scatters but big between-class scatters, the Fisher discrimination enhances the discrimination of the dictionary and sparse codes. We apply the proposed algorithm to four image datasets:AR, Brodats, JAFFE and AR gender. Experimental results show the effectiveness of LRFDDL. 2. Propose a dictionary learning algorithm based on label consistent and Fisher discrimination, which is called LCFDDL. We incorporate the label consistent term and the Fisher discrimination into the objective function to get a more discriminative dictionary and sparse codes. In order to enhance the discrimination of the dictionary, the label consistent term makes the label information of samples is associated with the label information of the atoms, in which the sample is represented by the dictionary atoms belonged to the same class as the sample; while the Fisher discrimination enhances the discrimination of the dictionary and sparse codes by minimizing the within-class scatters and maximizing the between-class scatters of the sparse codes. Due to the discrimination of the learned dictionary and sparse codes, we utilize the learned dictionary and sparse codes for classification to improve the classification performances. We apply the proposed algorithm to four image datasets:AR, Brodats, JAFFE and AR gender. Experimental results show the effectiveness of LCFDDL.3. We apply LRFDDL and LCFDDL to hyperspectral image classification. First, we get the texture features of the hyperspectral images by Gabor and fuse the texture features and spectral features of the hyperspectral images. Then we apply the proposed dictionary learning methods for hyperspectral image classification. We experiment on four hyperspectral images:Pavia University, Pavia University Center, Washington DC and Salina-A. Experimental results show that LRFDDL and LCFDDL can improve the performances of hyperspectral image classification and further verify the effectiveness of LRFDDL and LCFDDL.