Region Adaptive Mode Selection Algorithms for Image Super-Resolution

Image super-resolution is to generate high-resolution images from low-resolution ones. There are a variety of super-resolution methods. Some are suitable in smooth regions. Some perform well in edge regions. Some achieve good performance in regular texture regions (e.g. the methods that apply the patch with similar structure). In this paper, first, we improve the patch based methods and use the similarities of smoothness and direction to choose proper reference patches. We call it the smoothness similarity and edge-direction based interpolation (SSEDI) method. Then, since different methods are suitable for different regions, we propose an adaptive method to fuse different methods according to multi-scaled first and second order differences. Three methods are chosen for fusion: KMDCC, the cubic B-spline, and the proposed SSEDI, which have good performance in smooth, edge, and regular structure regions, respectively. The derived super-resolution method will perform well in all of the three parts of images.