Image super-resolution via multi-view information fusion networks

摘要

In recent years, convolution neural networks (CNN) have achieved substantial advantages for single image super resolution (SISR). However, as the depth and width of the networks increase, the model parameters and computation become more complicated. To solve these problems, we propose a lightweight network structure for super-resolution which adopts the multi-view information fusion strategy. The proposed network consists of a feature extraction block, N channel-fusion enhancement blocks (CFEBs) and an upscale block. In order to reduce model parameters and computation complexity, each CFEB adopts group convolution operation and channel-fusion strategy to fuse image information from different channels. In addition, we propose a global-relation attention block to enable interactions among groups and an enhancement block to better extract and fuse multi-scale feature information. Experimental results on public datasets demonstrate that both the proposed model and its lightweight version achieve superior performance than their state-of-the-art couterparts. Moreover, the proposed lightweight model sometimes even has better performance than some state-of-the-art heavy models. (C) 2020 Elsevier B.V. All rights reserved.