A Spatial Focal Modulation Network for Image Restoration

Deep learning methods based on CNNs and Transformers have attained notable achievements in image restoration (IR). However, CNNs have limited receptive fields, which can be a constraint in IR tasks for specific input content. The self-attention mechanism of Transformers is affected by quadratic computational complexity, making it particularly noticeable in IR tasks with high-resolution inputs. This paper proposes a Spatial Focal Modulation Network (SFMN) for IR, attempting to introduce focal modulation to replace the selfattention mechanism for modeling. Experiments on the public datasets Set5, Set14 and Urban100 show that SFMN achieves competitive results.