Hybrid Domain Attention Network for Efficient Super-Resolution

Abstract

Image SR reconstruction methods focus on recovering the lost details in the image, that is, high-frequency information, which exists in the region of edges and textures. Consequently, the low-frequency information of an image often requires few computational resources. At present, most of the recent CNN-based image SR reconstruction methods allocate computational resources uniformly and treat all features equally, which inevitably results in wasted computational resources and increased computational effort. However, the limited computational resources of mobile devices can hardly afford the expensive computational cost. This paper proposes a symmetric CNN (HDANet), which is based on the Transformer’s self-attention mechanism and uses symmetric convolution to capture the dependencies of image features in two dimensions, spatial and channel, respectively. Specifically, the spatial self-attention module identifies important regions in the image, and the channel self-attention module adaptively emphasizes important channels. The output of the two symmetric modules can be summed to further enhance the feature representation and selectively emphasize important feature information, which can enable the network architecture to precisely locate and bypass low-frequency information and reduce computational cost. Extensive experimental results on Set5, Set14, B100, and Urban100 datasets show that HDANet achieves advanced SR reconstruction performance while reducing computational complexity. HDANet reduces FLOPs by nearly 40% compared to the original model. ×2 SR reconstruction of images on the Set5 test set achieves a PSNR value of 37.94 dB.

Keywords

• super-resolution
• attention mechanism
• deep learning
• symmetry