Image Super-resolution Via Sparse Representation

The resolution of the image determines the richness that the image details are described, and how to improve the resolution of existing image is a hot topic in the image processing technology currently. In order to improve image resolution on the basis of the existing image without increasing the hardware cost, researchers proposed super-resolution reconstruction technique. Super-resolution reconstruction technology is based on software of image processing technology, the technology can be divided into three approaches according to different processing methods: interpolation reconstruction, reconstruction of constraints and learning-based.Interpolation reconstruction method is simple and easy to implement, but the reconstruction of points by this way is the sum of around points with their weights, the image information entropy does not increase, and there is no enhancement in image details. Constraint reconstruction is a method that uses multiple frame image fusion to reconstruct a new high resolution image, which has an increase in information entropy compared with the single frame and low resolution image before reconstruction. But under the condition of larger ratio of image reconstruction and the less low resolution images, the effect of reconstruction declined seriously.Learning-based method is a new reconstruction technique in recent years, this method selects the images that contain the same information with the waiting for reconstruction image as the training sample, and obtains prior knowledge from a large number of training samples as the basis of super-resolution reconstruction, during the reconstruction using the knowledge gained in the reconstruction of the image, and the prior knowledge of the image itself on the basis of the waiting for reconstruction image, the high resolution image can be reconstructed under the condition of the input image with a single frame, and overcome the limitations of the constraint reconstruction method in aspects of improving the reconstruction of multiple. The method based on learning mainly includes: Freeman’s based on the sample method, Chang’s embedding method and Yang’s based on sparse representation method.This paper mainly studies the Yang’s based on sparse representation method, including two processes of dictionary training and image reconstruction. In dictionary training, adopting the method of high joint training and low resolution in the process of image reconstruction, solving the input low resolution block coefficient of sparse representation under the low resolution of a dictionary firstly, and then using the coefficient of dictionary and high resolution to reconstruct the corresponding high resolution block.This thesis makes an improvement on the method of Yang’s, this method realizes that the reconstruction constraint conditions are consistent with the process of the dictionary training, jointing high and low resolution block to calculate sparse representation coefficient, that is, using the first reconstruction high resolution blocks and the low resolution input blocks under the joint dictionary to make a sparse decomposition again. This method can be repeated iteration. In the experiments, the method in this paper on the binomial parameter values of visual effect, PSNR and SSIM are better than Yang’s.This paper puts super-resolution reconstruction technology into the characters of the image recognition, three kinds of methods are put forward according to the characteristics of the image design, the experiments show that super-resolution reconstruction technique can improve the character image resolution, in order to improve the recognition rate, and the experiment result shows a feasible solutions of the super-resolution reconstruction in character image recognition applications. Finally, the cross reconstruction contrast experiment between different samples shows the dictionary who obtained by the learning method is adaptive, but the image reconstruction changes little in the vision, it is a resistance for transferring the research results to practical application, it also may be momentum for further study.