Deep Learning Based Single Image Deblurring

The image taken by camera often appear blurry artifacts due to camera shake,object motion,and scene depth changes.Such artifacts seriously affects the quality of the image and the accuracy of other computer vision tasks,such as face recognition and target detection.Single image deblurring aims to reconstruct a clear image from a blurry image,which is of great significance to photography and computer vision.Traditional deblurring methods are timeconsuming and difficult to handle blurry images of complex scenes.Although the existing deep learning methods can effectively reconstruct clear images,they have certain deficiencies to reconstruct the high-frequency information of the blurry image,resulting in the reconstructed image being too smooth,and they are limited by the partial blur in the image,leading to poor deblurring effect.Based on deep learning technology,this paper made the following contributions.First,we propose a deblurring method based on the high-frequency prior,aiming to strengthen the convolutional neural network's ability to process high-frequency information of the blurry image.We propose a novel image deblurring framework to focus on the reconstruction of high-frequency information,which consists of two main subnetworks: a highfrequency reconstruction subnetwork and a multi-scale grid subnetwork.The high-frequency reconstruction sub-network extracts the clear high-frequency features from the blurry image,and then use it to guide the multi-scale grid sub-network to restore clear image.In addition,in order to make better use of high-frequency features,we designed two different guidance methods.Under the guidance of high-frequency features,the deblurred image recovered by our method has clearer details.Second,aiming at the problem of local blur inconsistency,we introduce a novel quadtree convolution module and a novel local spatial attention module into our deblurring network.The quadtree convolution module divides the image features into independent small pieces,and then extracts and reconstructs its local features separately.The local spatial attention module enhances the network's ability to perceive blurry areas by calculating the correlation between each pixel and its surrounding pixels.With the cooperation of the quadtree convolution module and the local spatial attention module,our method can better deal with the local blur in the image.Extensive experiments on synthetic datasets and real-world blurry images demonstrate that our method achieves competitive results.Methods proposed in this paper has important reference significance for both convolutional neural network and image deblurring.