Research On Motion Blurry Image Restoration Based On Generative Adversarial Networks

Motion blur is one of the most common image blur types.Motion blur is caused by the relative motion between the camera and the object being photographed.It will affect image quality and reduce the accuracy of many algorithms.Due to the widespread existence of motion blur and the fact that the deblurring algorithm is still unable to restore the image well,motion blur removal is still an popular research orientation in the field of image processing.In recent years,some achievements have been made in the research of blurry image restoration using deep learning.In this paper,the main research of motion blurry image restoration algorithm base on generative adversarial networks are as follows:(1)The development and problems of traditional blurry image restoration algorithms and blurry image restoration algorithms based on deep learning are discussed.The principle of deep learning and motion blurry image restoration algorithm base on generative adversarial network is described in detail.(2)A motion blurry image restoration algorithm base on enhanced generative adversarial networks is proposed.By using resize convolution instead of deconvolution as upsampling strategy on image,the checkboard artifacts caused by uneven overlap is solved effectively;By using the characteristics of squeeze and excitation network which can be used to extract the weight of the feature channel,the squeeze and excitation network and residual network are combined to effectively improve the restoration effect of the motion blurry image.The multi-dimensional loss function enables the final loss function of the network to compare the image in pixel level,high-dimensional feature level and gradient domain at the same time,and improve the restoration effect of motion blurry image.(3)A motion blurry video restoration algorithm base on multi-frame information and 3D deformable convolutional networks is proposed.Since video contains a large amount of inter-frame information in addition to the in-frame information.There is anassociation between the front and back frames.When the input is video,the enhanced generative adversarial networks cannot use multi-frame image to help the motion blurry image recovery.Therefore,we can takes the current frame and its front and rear two frames as input,the output is the superposition of a feature map generated by the network and the current frame.VDGAN can utillize multi-frame information to improve motion blurry image restoration.In addition,3D deformable convolution is proposed.Through the three-dimensional deformation of the sampling grid,3D deformable convolution can effectively extract the spatio-temporal information in video,and can better model the feature deformation of video caused by motion blur.3D-VDGAN improves the performance of 2D standard convolution.(4)The proposed model is lightened.The superior performance of deep learning is largely dependent on the large number of network parameters,but is also brings many problems,such as large demand for computing resources,long computing time,large space and high computing cost requirement.In order to solve these problems,the model is lightened to reduce the number of network parameters and improve the network efficiency.