G-HAPNet: A Novel Structure for Single Image Super-Resolution

Abstract

Recent investigations on single image super-resolution (SISR) have progressed with the development of deep convolutional neural networks (DCNNs). However, increasingly complex network designs cause huge computational budgets. Therefore, a more efficient structure for SISR task is desirable. In this report, we propose a novel structure, called G-HAPNet. Specifically, the group-hierarchical atrous pyramid block (G-HAPB) is built to package as a general block for deeper network constitution. Firstly, the original features are expanded and grouped in channel. Then, a atrous pyramid is constructed to extract multi-scale features from corresponding channels. Besides, we introduce hierarchical grouping aggregation (HGA) which includes forward aggregation and backward aggregation by skip connections so that we can achieve hierarchical fusion and information guidance among multi-scale features. Extensive experiments demonstrate that with the same level depth and computational budgets, our proposed G-HAPNet has better performance than state-of-the-art methods on both synthetic datasets and real-world dataset, which indicates our G-HAPNet is a more efficient and practical structure for SISR.

This work was supported in part by the National Natural Science Foundation of China under Grants 61571382, 81671766, 61571005, 81671674, 61671309, 61971369 and U1605252, in part by the Fundamental Research Funds for the Central Universities under Grants 20720160075 and 20720180059, in part by the CCF-Tencent open fund, and the Natural Science Foundation of Fujian Province of China (No. 2017J01126).