Application Research On Image Sparse Representation And Image Super Resolution

With the existing imaging devices and observation instead of high-cost chips and imaging systems, image super-resolution uses the low-cost software methods to reconstruct high-resolution images and is a hot spot in the image research area. Image super-resolution, based on less image data is a typical ill-posed problem. To make the solution unique and stable, the image prior knowledge has to be exploited.Sparse representation theory is wildly studied in pattern recognition and image processing and is also an efficient learning-based method in single image super-resolution. Based on the sparse representation theory, this thesis focuses on redundant dictionary learning and sparse representation methods applied in single image super-resolution. Sparse representation based super-resolution involves three steps: first, off-line training of the redundant dictionary pair to build high- and low-frequency information; next, on-line calculation of the sparse coefficients of the low resolution image in terms of the low-resolution dictionary; finally, reconstruction of the high-resolution image from the coefficients and high-resolution dictionary.This thesis addresses one of the challenges in how to construct efficient redundant dictionary. By analyzing traditional and joint learning dictionary, this thesis proposes an improved joint dictionary which contains the graph regularized information. The dictionary introduces the graph regularized similarity in sparse coefficients to enhance constraints between patches. When applied in single SR, simulation results show that the graph regularized dictionary provides more image information, produces better high-resolution images both in subjective visually and quantitative evaluations.This thesis also proposes an efficient sparse representation method for single image super-resolution called feature sign method. By guessing the sign of sparse coefficients, the complicated 1-norm question can be changed to a QP question. Simulation results demonstrate the advantage of the proposed scheme over existing schemes. While output images from the Bicubic have edge blur, images from OMP have badly jagged artifacts, those from LARS have some blocky effects, and PCG reconstructions are too time wasting, the images from feature sign have more distinct visual details. RMSE and SSIM of reconstructed images illustrate the good quality of proposed method over the compared ones.