Learning Omni-Frequency Region-adaptive Representations for Real Image Super-Resolution

• DOI: 10.1609/aaai.v35i3.16293
• Publication Date: 2022-09-08T18:10:11Z
• Authors: Xin Li; Xin Jin; Tao Yu; Simeng Sun; Yingxue Pang; Zhizheng Zhang; Zhibo Chen
• Keywords: Feature (linguistics); Upsampling; Convolution (computer science); Spatial frequency

Traditional single image super-resolution (SISR) methods that focus on solving and uniform degradation (i.e., bicubic down-sampling), typically suffer from poor performance when applied into real-world low-resolution (LR) images due to the complicated realistic degradations. The key this more challenging real (RealSR) problem lies in learning feature representations are both informative content-aware. In paper, we propose a Omni-frequency Region-adaptive Network (OR-Net) address challenges, here call features of all low, middle high frequencies omni-frequency features. Specifically, start frequency perspective design Frequency Decomposition (FD) module separate different components comprehensively compensate information lost for LR image. Then, considering regions have lost, further Aggregation (RFA) by leveraging dynamic convolution spatial attention adaptively restore regions. extensive experiments endorse high-efficient, effective, scenario-agnostic nature our OR-Net RealSR.