ASSR: A Lightweight Super Resolution Network with Aggregative Structure

Abstract

With the rise of deep learning, the effect of Super- Resolution task has been greatly improved. Many studies have shown that the residual structure and the dense structure have a significant effect on the Super-Resolution task. However, there are three shortages of these kinds of methods: huge amount of parameters, heavy computaion, and bad reconstruction effect of details and textures. In this paper, we propose six design guidelines for designing lightweight Super-Resolution network which can not only alleviate the number of parameters and complicated calculation problems effectively, but also improve the visual effects of the reconstruction results. Following the proposed guidelines, we design two lightweight Super-Resolution networks: ASSR and ASSR-s. By comparing with the state-ofthe- art methods, we prove the effectiveness of the proposed designing guidelines. Our ASSR and ASSR-s are five times faster with comparable and even better PSNR and visual effects with the state-of-the-art models. In the AIM 2019 Constrained Super-Resolution Challenge, the proposed ASSR won the second place of track 1 and the third place of track 2.