AU-Shaped Fourier Unit Transformer for Single Image Deblurring

There are many causes of image blurring, including optical, atmospheric, artificial and technical factors, etc. It is important to deblur images in daily production life. To achieve better results, different methods are needed to deal with different causes of blurring. From the technical aspect, blurred image processing methods are divided into three main categories, namely image enhancement, image restoration and super-resolution reconstruction. Overall, although blurred image processing algorithms have achieved a very wide range of applications, but image algorithms have their own limitations after all, we can not pin all problems on the image algorithm, for different kinds of blurred problems, to be treated differently. For blurring or image quality degradation caused by lens out of focus, dust obscuring, line aging, camera failure, etc., with the help of video diagnostic system, must be repaired in time to solve the problem at the source. For low light and other priority to choose day and night type high-sensitivity cameras, for rain and fog, motion and pre-sampling caused by image quality degradation, you can use the “video enhancement server” contains a variety of blur image processing algorithms to improve image quality. The purpose of single-image deblurring is to restore blurry images to sharp images and recover the texture detail features of the image. The Transformer model is able to capture the correlation of the blurry pixels over long distances, which has significant performance on image deblurring tasks, but its computational complexity grows quadratically with increasing spatial resolution. In addition, some methods use convolution to reduce image resolution, but it leads to information loss. In this paper, we propose an effective Transformer model, named Fourier Unit Transformer (FUformer), based on Fourier transform constructing the shallow features extraction module and the self-attention module, which can reduce computational complexity and fuse image local and global features to recover image texture details. Experimental results show that the proposed method performs well on the deblurring task. Code will be available at https://github.com/zox-iii/FUformer.