Fourier transform-based u-shaped network for single image motion deblrring

Image deblurring is one of the fundamental tasks in image processing tasks, which can provide the necessary support for advanced tasks such as image recognition. In this paper, we propose a new deblurring model, named Efftformer model, which is specialized in the blurring elimination of motion blur. The model focuses on the recovery of detail information and edge information to provide more effective image information and better basic support for the realization of advanced tasks. In Efftformer model, firstly, we introduce a frequency domain based ReLU residual stream, which allows network to learn blur kernel level information for better restoration of original image. Secondly, we propose a cross-connection channel attention module (CCAM) to explore an effective fusion approach in multiple scales adaptively, which helps decoders to restore original image well by aggregating the semantic information in different scales. Considering the effectiveness of edge information in image recognition tasks, we enhanced the edge information in recovered image by performing a Sobel filter as well as an auxiliary edge loss function. We conducted experiments on different motion blur datasets and compared them with state-of-the-art algorithms. The experimental results show that Efftformer model proposed in this paper achieves comparable even superior performance to the state-of-the-art algorithms.