Single Image Super-Resolution with Attention-based Densely Connected Module

Abstract Benefited from the abundant features provided by the dense connection block, the densely connected-based super-resolution network has achieved superior performance in the single image super-resolution (SISR) task. However, the abundant features also introduce redundant and conflicting information, resulting in longer training time and unsatisfied image reconstruction results. To solve this problem, we propose an attention-based densely connected module (DAM). DAM consists of two parts: channel attention module (CAM) and dense connection block (DB). CAM is placed at the front of each DB and gives different weights of each channel from received features for suppressing redundant responses. Based on DAM, we propose an Attention-based Densely Connected Network (ADSRNet) for SISR, and explore the effectiveness of DAM on other densely connected-based super-resolution networks. Extensive experiments are performed on commonly-used super-resolution benchmarks. Qualitative and quantitative results demonstrate the effectiveness of our method.