Lightweight Dynamic Hybrid Attention Network for Single Image Super-Resolution

Single image super-resolution (SISR) is of great importance and has found wide applications. In recent years, convolutional neural network (CNN) has been implemented to SISR and achieved great successes. However, many existing CNN-based SR networks improve the performance by stacking a large number of blocks, resulting in more parameters and higher computation cost. To solve this problem, we propose a lightweight dynamic hybrid attention network (DHAN) which is made up of chained dynamic hybrid attention blocks (DHABs). A DHAB is mainly composed of a non-attention branch and a hybrid attention branch based on 3D convolution kernels, and can combine channel attention and spatial attention to extract powerful representations of inter-channel and intra-channel feature response. Furthermore, we design a dynamic weight module of DHAB, which generates two sum-to-one weights according to the input features of each block and dynamically adjusts the contribution of the hybrid attention branch and the non-attention branch. Based on the hybrid attention branch and the dynamic weight module, DHAN can not only efficiently reduce the number of multiply-add operations (Multi-Adds) and the number of parameters, but also accurately recover the textural details of the high-resolution image from the low-resolution feature map. Experiments on benchmark datasets are done to demonstrate that the proposed DHAN achieves excellent performance compared to some state-of-the-art networks of similar sizes.