Motion Deblur with Non-Local Attention Network

Abstract

Motion blur, which degrades image quality significantly, is a common and huge obstacle in many other image processing applications. And deep learning has been used in several fields of image processing in recent years. In this paper, we present an efficient motion deblur network based on the Non-Local Attention Network. This network can deblur an image blurred by motion blindly without any prior knowledge. Our network follows the encoder-decoder structure, and a residual network module consisting of multiple residual networks is added to both the encoder and the decoder to extract the depth features of the input feature maps. Local and non-local attention modules built according to the residual network idea are also added to the network, which in turn improves the network's ability to capture long-term dependencies and allows us to build deeper networks to improve the expressiveness of the network. Experiments have shown that our method achieves quantitatively and visually comparable or better results than current leading methods.