SRFLOW-Attention: Super Resolution with Multi-Head Attention for Effective Use of Low-Resolution Information

Abstract

This paper proposes normalizing flow-based image super-resolution techniques using attention modules. In the proposed method, the features of the low-resolution images are extracted using a Swin Transformer. Furthermore, multi-head attention in the flow layers makes effective use of the feature maps. This architecture enables the efficient injection of low-resolution image features extracted by the transformer into the flow layer. Experimental results at x4 magnifications showed that the proposed method achieved state-of-the-art performance for quantitative metrics and visual quality among single-loss architectures. © 2024 Institute of Electrical Engineers of Japan and Wiley Periodicals LLC.