SCAN: Spatial Color Attention Networks for Real Single Image Super-Resolution

Conceptually similar to adaptation in model-based approaches, attention has received increasing more attention in deep learning recently. As a tool to reallocate limited computational resources based on the importance of informative components, attention mechanism has found successful applications in both high-level and low-level vision tasks which includes channel attention, spatial attention, non-local attention and etc. However, to the best of our knowledge, attention mechanism has not been studied for the R, G, B channels of color images in the open literature. In this paper, we propose a spatial color attention networks (SCAN) designed to jointly exploit the spatial and spectral dependency within color images. More specifically, we present a spatial color attention module that calibrates important color information for individual color components from output feature maps of residual groups. When compared against previous state-of-the-art method Residual Channel Attention Networks (RCAN), SCAN has achieved superior performance in terms of both subjective and objective qualities on the dataset provided by NTIRE2019 real single image super-resolution challenge.