Dual-Domain Attention for Image Deblurring

As a long-standing and challenging task, image deblurring aims to reconstruct the latent sharp from its degraded counterpart. In this study, bridge gaps between degraded/sharp pairs in spatial frequency domains simultaneously, we develop dual-domain attention mechanism for deblurring. Self-attention is widely used vision tasks, however, due quadratic complexity, it not applicable with high-resolution images. To alleviate issue, propose novel module by implementing self-attention style of dynamic group convolution integrating information local region, enhancing representation learning capability reducing computational burden. Regarding domain learning, many frequency-based approaches either treat spectrum as whole or decompose components complicated manner. work, devise compactly decouple into distinct parts accentuate informative part extremely lightweight learnable parameters. Finally, incorporate modules U-shaped network. Extensive comparisons prior arts on common benchmarks show that our model, named Dual-domain Attention Network (DDANet), obtains comparable results significantly improved inference speed.