Real-Time Texture Synthesis via Aggregation of Internal and External Patch Candidates

In practical image display systems, the image quality of magnified videos is considered very critical. Although many image super-resolution algorithms based on deep learning using external database were proposed for the upscaling of low-resolution contents, they still do not provide excellent performance in high-resolution displays. This may be due to several troublesome issues such as the lack of analysis of high-frequency texture components, the adopted edge-oriented loss function, and the use of external database different from the test image. In this study, we focus on the difficult issues and propose a real-time texture synthesis algorithm, based on aggregation of internal and external patch candidates. Specifically, to improve the image fineness while keeping texture naturalness, merged and rearranged distant pixels in the input image are combined with external examples obtained from a hardware-friendly technique specially redesigned. Experimental results show the proposed texture synthesis algorithm outperforms existing state-of-the-art algorithms.