Single image super resolution with wide activation for mobile devices

Abstract

Single Image Super Resolution (SISR) revolves around the task of reconstructing a high-resolution image from a single low-resolution image. Numerous applications of SISR range from surveillance & security, medical imaging to photographic utilities. Although there are ample SISR solutions, especially those which are deployed as cloud services, there’s a scarcity of effective on-device mobile SISR solutions. Even the existing solutions are mostly limited to high end mobile devices and most of the time limited by device architecture. An effective SISR solution which can run on any mobile device would be extremely helpful to the community in this context and can help gain a number of benefits in an edge-computing point of view, including storage and transfer optimization for image content. This research primarily focuses on creating such a solution, specifically focusing on usage of on-device Wide Attention Networks (WDSR) for SISR. In addition, a performance comparison will be done with other CNN based models.