Recently, various deep-neural-network (DNN)-based approaches have beenproposed for single-image super-resolution (SISR). Despite their promisingresults on major structure regions such as edges and lines, they still sufferfrom limited performance on texture regions that consist of very complex andfine patterns. This is because, during the acquisition of a low-resolution (LR)image via down-sampling, these regions lose most of the high frequencyinformation necessary to represent the texture details. In this paper, wepresent a novel texture enhancement framework for SISR to effectively improvethe spatial resolution in the texture regions as well as edges and lines. Wecall our method, high-resolution (HR) style transfer algorithm. Our frameworkconsists of three steps: (i) generate an initial HR image from an interpolatedLR image via an SISR algorithm, (ii) generate an HR style image from theinitial HR image via down-scaling and tiling, and (iii) combine the HR styleimage with the initial HR image via a customized style transfer algorithm.Here, the HR style image is obtained by down-scaling the initial HR image andthen repetitively tiling it into an image of the same size as the HR image.This down-scaling and tiling process comes from the idea that texture regionsare often composed of small regions that similar in appearance albeit sometimesdifferent in scale. This process creates an HR style image that is rich indetails, which can be used to restore high-frequency texture details back intothe initial HR image via the style transfer algorithm. Experimental results ona number of texture datasets show that our proposed HR style transfer algorithmprovides more visually pleasing results compared with competitive methods.
展开▼
