A Compact Network Using Attention Mechanism for Image Super Resolution

With the emergence of the 2K/4K Ultra High-Definition (UHD) videos super-resolution, single image super resolution based on convolutional neural networks (CNNs) is a critical method to transfer Standard Definition (SD) videos to High Definition (HD) videos. However, the state-of-the-arts methods rely on a deeper or wider network, which poses a challenge in computational complexity and memory consumption. To address this problem, we proposed a compact super-resolution framework using refined features. In addition, an attention channel sorted split (ACSS) module was proposed to significantly improve the performance of super-resolution (SR). Especially, the ACSS module can make the network pay more attention to transfer features from coarse to fine. The experimental results show that our proposed method is efficient and effective.