Natural Image Group Based On Sparse Representation Of Super-resolution Algorithm

The so-called image super-resolution is defined by given one or a series of low resolution images, combined with the known prior knowledge to get the high resolution image of the same scene. In recent years, with the development of pattern recognition and computer vision, image super-resolution method based on learning can get better results than traditional methods, so has been widely studied. The super-resolution reconstruction method based on sparse representation is one of them and has proved its validity in practice. Now, the group sparse representation, which is based on the sparse representation technique, has been further promote the development of the reconstruction technology in the aspect of theoretical research and practical application, it can reduce the number of iterations and the reconstruction time under a certain sparse degree. This paper research the group sparse representation in the application of image super-resolution. Aiming at the limitations in the process of group sparse, we focus on analyzes and introduces the prior knowledge as the regularization item dictionary during training process, meanwhile obtain the structured dictionary and apply this method to the natural image super-resolution reconstruction and image denoising. In this paper, the main research work is as follows:First, although image super-resolution method based on group sparse representation can achieved super-resolution image reconstruction, but because of the high requirements of its similarity to the training and testing samples, the using scope is greatly restricted. Based on the analysis of the before relevant literature, we achieve the natural image super-resolution based on the group sparse representation.Secondly, this paper figure out a structured dictionary training method based on dictionary clustering and regularization item is presented in this. The main idea is add the prior information of the trained dictionary to the training process during the dictionary updates, resulting in a structured dictionary. Because of using this method, the dictionary that can response more sample information contains a large number of prior information than the traditional dictionary, which has very good effect in the natural image super-resolution.Moreover, for improving the structural dictionary based on the dictionary clustering and regularization, the training method based on the fisher regularization item is introduced. From the point that the previous dictionary learning method just impose constraints on atoms which is in the different clusters, without considering the situation of the atoms in the same cluster, the new improved dictionary training method which integrated into the deformation of the Fisher discriminant information, make the using of the dictionary structure in the process of the image reconstruction based on group sparse more apparent and have stronger ability for natural images reconstruction.Finally, applying the trained structured dictionary to image noise reduction, we realized the natural image noise reduction based on group sparse representation.