Edgeformer: Edge-Enhanced Transformer for High-Quality Image Deblurring
Yuan Zou, Yinyao Ma · 2023
Transformers in image deblurring have achieved significant progress recently under their outstanding ability to model the long-range information of features. However, the modeling in the blurry image is challenging due to the ambiguous correlation caused by the diversity of the blur degree. Whereas the modeling works well in clear regions, which can be represented by edges or textures. Therefore, we explore the effectiveness of edges and textures in image deblurring and present an effective and efficient edge-enhanced Transformer, called Edgeformer, to acquire high-quality images. Specifically, we develop an efficient edge-enhanced attention (EA) that focuses on the sharp areas when computing attention maps. Furthermore, we introduce an edge-enhanced feed-forward network (EFFN) that could discriminate edge and texture features preserved for latent clear image deblurring. Experimental results show that the proposed method performs favorably against the state-of-the-art methods.