Optimized Image Up-Scaling from Learning Selective Similarity

Abstract

Remarkable up-scaling results are obtained from local self-examples (LSE) [1] at low cost. However, fine-detailed and cluttered regions are not reproduced realistically due to inappropriate addition of high frequency, and thus appear somewhat faceted and slightly distorted at edges. In this paper, an optimized algorithm that reduces these artifacts is proposed. A selective search is applied in the original restricted searching area. Mismatches can be avoided when patches are with extremely random details. Meanwhile we extend the local self-similarity on natural images to sub-pixel level which makes the assumption of local self-similarity more suitable especially for fine edge areas. It corrects the slightly misalignments when scaled with small factor. The proposed algorithm remains simple, and generates clear and believable textures compared with the original LSE and other mainstream methods.