Super-resolution Image Reconstruction Algorithm Based On Sparse Representation

With the continuous development of technology, people have high requirements for image resolution. How to improve the image resolution has become a hot topic in the field of image processing in recent years. Image super-resolution(SR) reconstruction is that a process reconstructs one or multiple low-resolution images of the same scene as a high-resolution image. SR reconstruction technique can effectively improve image resolution. It can apply in many fields such as medicine diagnostics, military and video surveillance etc.At present, super-resolution reconstruction algorithms can be divided into three classes: based on interpolation algorithm, learning algorithms and reconstruction algorithms. Reconstruction algorithms and learning algorithm are widely used. This thesis focuses on the sparse representation algorithm based learning for single image problem. The main works areas following aspects:1. Ill-posed problem of super-resolution reconstruction can use regularization method to solve, but the parameter of traditional regularization method is a fixed value. So image edge and texture detail information is not preserved well for reconstructed image. A image adaptive regularization super-resolution algorithm based on L1 norm and BTV is proposed. This algorithm uses adaptive control regularization parameters to control the objective function data fidelity term and regularization terms of proportion to solve this problem. The simulation results show that the proposed algorithm has better reconstructed image quality than the traditional regularization algorithm.2. In order to better retain the image edge and texture characteristics in super-resolution reconstruction process, a reconstruction algorithm based on sparse representation regularized super-resolution is proposed in this paper. Firstly, K-SVD method is used to get a single over-complete dictionary, and then the joint dictionary construction method is combined to obtain high and low resolution of the dictionary, and finally the sparse representation coefficient is solved by optimization algorithm and combined with high-resolution dictionary to get high resolution images.3. Due to the sparse representation algorithm requires a lot of training samples to train dictionary, its calculated amount is larger and reduces learning efficiency of dictionary. To solve this problem, a super-resolution image reconstruction algorithm based on clustering is proposed. First clustering classification is done for input image samples, and principal component analysis(PCA) is used to reduce calculated amount for each sub-category, each type of training samples is trained and learned for the corresponding sub-dictionary. The high and low resolution dictionaries are obtained. Finally, low-resolution image is reconstructed by using the product of the dictionary of high-resolution images block and sparse representation. Compared with traditional interpolation method and Elad method, the proposed algorithm, the proposed is more feasible and effective in simulation experiments.