Research On Image Super-Resolution Based On Sparse Constraint

Image super-resolution is to reconstruct a high resolution image from one or more low resolution images of the same scene by using some certain prior knowledge. This technology has a good prospect for its device independence and encouraging results.This thesis first briefly introduces some classical algorithms and some theories about image sparse and redundant representation model, and then focuses on the research of image super-resolution based on image non-local redundancy sparsity and dictionary sparse representation. The three contributions presented in this thesis can be concluded as follows.1. An improved non-local iterative back-projection algorithm (NLIBP) is proposed. It can reduce the computational complexity and avoid the over-correction for some interpolated pixels by adaptively controlling the number of pixels involved in the non-local modifying process and optimizing the similarity calculation between pixels. Experimental results show that this improved algorithm reduces the reconstruction complexity and generates images with higher subjective and objective quality.2. An adaptive fast reconstruction method based on the K-Means clustering is presented to reduce the reconstruction computation complexity consumed by the algorithm based on the dictionary sparse constraint model. This presented algorithm reduces its complexity from two aspects. (1) Reduce the dictionary size for each image patch in the learning process by classifying the sampled raw patches in the dictionary training process. (2) Adaptively select the reconstruction algorithm according to the features existed in each patch. Experimental results show that this proposed fast reconstruction method takes much less time while generating images equivalent to the original algorithm.3. A global post-processing method guided by non-local similarity structure is proposed to improve the reconstruction quality gained by the algorithm based on dictionary sparse constraint model. It makes full use of the edges and the non-local similarity structure existed in the low resolution image and combines the non-local means (NLM) denoising algorithm and the improved NLIBP algorithm successfully. Experimental results show that this proposed post-processing method can effectively improve image edges and its overall smoothness, generating images with higher subjective and objective quality.