研究目的

The paper aims to solve two key issues in image super-resolution (SR): obtaining and using valid image prior information, and solving the sparse coefficients more efficiently. It proposes a novel super-resolution reconstruction method via self-similarity learning and conformal sparse representation to improve the visual performance and algorithm stability.

研究成果

The proposed super-resolution reconstruction method via self-similarity learning and conformal sparse representation exploits the self-similarity of the image and combines the local geometric angle invariance of the similar patches and their corresponding coefficients. It retains the local structural information of the image and captures the global structure of the data with low-rank constraints. Experiments show that the method has the highest average values of PSNR and SSIM with the best visual performance and clearer texture, although it is not the fastest due to the conformal characteristics.

研究不足

The proposed method cannot guarantee convergence to global optimum, although experience shows that the algorithm is convergent. The method does not run very fast due to the extra computation of the conformal relationship, especially for big datasets.