Multi-scale Visual Aggregation Residual Network for Super-Resolution

Abstract

Deep neural networks with residual network structure have achieved great progress in the fields of super-resolution and have greatly improved the performance of super-resolution. However, in order to obtain higher reconstruction accuracy, simply increasing the depth of the network will aggravate the complexity of the network. In this paper, an efficient multi-scale visual aggregation residual network (MVARN) structure is developed for exploiting abundant image detail features, which mimics the visual aggregation mechanism of human eyes. Within the framework of the residual network, we design convolution kernels of different sizes to exploit image detail features by the manner of gradually increasing the range of receptive fields. We show that by utilizing the residual network framework with the multi-scale convolution (MSC) kernels, the image detail feature is capable of being exploited effectively. The feature extracted from different scales can complement and fuse each other to obtain the desired image information. Experimental results show that our proposed method achieves visually pleasing results and is superior to some of the state-of-the-art methods.