Image Super-resolution Based On Self-similarity and Various Patch Size

A new single image super-resolution method based on self-similarity across different scales and pyramid model is proposed. In order to enrich the diversity of the training patches but not increase the computational complexity, we rotate the low resolution input image by a certain angle from 0 degrees to 90 degrees and down-sample them into 2 layers pyramid model respectively. However, most self-similarity super-resolution algorithms was carried out by the fixed size of patch. So, in this paper we observe the effect of patch size using the various patch size then pick out the most appropriate patch size. During the mapping process, we use the Fast Library for Approximate Nearest Neighbors (FLANN) to search the corresponding nine closest patches in high-frequency pyramid then carry out Gaussian weighted (SSD), which can avoid the occasionality and mismatch by using the nearest neighbor strategy. Finally, the local constraint and the iterative back projection algorithm are adopted to optimize the reconstructed image. Experimental results validate that the algorithm is better than the traditional algorithm in visual effects and computational complexity.