Two-Stream Sparse Network for Accurate Image Super-Resolution

Abstract

Deep convolutional neural networks (CNN) have achieved significant success in single image super-resolution (SISR) task. Though previous works have proved that deeper networks yield better performance, simply stacking more layers to deepen networks ignores the complementarity between different feature levels and largely increases the computation cost, and leads to limited image reconstruction quality. In this paper, we aim to separately extract shallow and deep features to enforce their respective contribution to the recovering of the super-resolution image. We propose a Two-Stream Sparse Network (TSSN) in which we design sparse residual block (SRB) to efficiently extract local features and explicitly learn shallow and deep features via two-stream architecture. Besides, we apply attention mechanism to effectively aggregate shallow and deep features. Extensive experiments demonstrate that the proposed method outperforms most of existing state-of-the-art methods on benchmark datasets in both quantitative metric and visual quality.