A Super-Resolution Video Player Based on GPU Accelerated Upscaling from Local Self-Examples

Many super-resolution algorithms have been proposed for upscaling static images, yet upscaling video footages in real-time with descent quality remains a challenging problem. In this work, we started out developing a super-resolution video player based on upscaling from local self-examples[Freedman and Fattal 2010], for the approach is essentially suitable for applications where maintaining temporal coherences between frames is critical. The algorithm takes advantage of the fact that a natural image patch is similar to some of its neighboring patches, and thus is redundant within its locality. Therefore, it utilizes neighboring patches, namely, selfexamples, as hints on the high-frequency detail that is lost in interpolative upscaling. High two-dimensional locality in image accessing pattern makes this algorithm an excellent candidate to be parallelized and implemented on commodity GPUs. With CUDA, we achieved a 75-times acceleration and a frame rate of about 20 fps for 1.5-times on-line upscaling of video of VCD quality. CR Categories: I.3.1 [Computer Graphics]: Hardware Architecture—Graphics Processors; I.3.3 [Computer Graphics]: Picture/Image Generation—Viewing Algorithms; I.4.3 [Image Processing and Computer Vision]: Enhancement—Sharpening and Deblurring;