Efficient Attention Fusion Feature Extraction Network for Image Super-Resolution.

Tuoran Wang,Na Cheng,Shijia Ding,Hongyu Wang

DOI: https://doi.org/10.1145/3613330.3613336

2023-01-01

Abstract:Many lightweight approaches for image super-resolution are currently unsatisfactory in their performance. To address this issue, we propose an efficient attention fusion feature extraction network (EAFFEN) for lightweight image super-resolution model. Our method is comprised of multiple efficient attention fusion feature extraction blocks (EAFFEB), each of which performs aggregation operations of deep feature extraction, enhancement, refinement, and redistribution of feature information. In each block, we employ the progressive residual feature extraction module (PRFEM) to extract deep information of image. Then the channel contrastive-aware attention module (CCAAM) is used to enhance the feature information. And finally, we use the channel-spatial joint attention module (CSJAM) to jointly redistribute the feature information from the previous layer and the enhanced deep feature information to collect various refined feature information. These three modules are concatenated in a cascade manner to form an efficient attention fusion feature extraction block. Our method can enable the performance while utilizing a small number of parameters. We evaluate our method on popular benchmark datasets and achieve superior performance compared to existing methods. For example, our EAFFEN achieves the PSNR of 30.50 dB and the SSIM of 0.9087 on Manga109 ×4 dataset.