Research On Blind Restoration Of Spatially Varying Blurry Image Based On Dynamic Filter Network

Motion blur is the most common blur in the daily image,which seriously affects the normal use of digital images.Motion blur images are usually divided into two types:spatially invariant blur images and spatially varying blur images.Each pixel in the spatially invariant blurry image has the same blur type and degree,that is,it has the same blur kernel.The blur types and degree of each pixel in the spatially varying blur image are not the same,that is,the types or size of blur kernels maybe different.For spatially invariant blurry image restoration,accurate estimation of global blur kernel is the key to solve the problem.However,it is extremely complicated and impractical to estimate the specific blur kernel for each blur pixel in a spatially varying blurry image.In recent years,by automatically learning the nonlinear mapping from blurry images to clear images,convolutional neural network can restore blurry images end-to-end without estimating blurry kernels.However,there are still some shortcomings in deep learning-based motion blurry image restoration methods.On the one hand,the weight of the convolutional kernel in the convolutional neural network is spatially invariant,so most existing networks use the convolution kernel with the same weight to filter in different spatial positions when restoring blurry images,ignoring the differences between different blurry pixels.On the other hand,most existing networks have a limited receptive field,which limits their ability to capture the global context relationship of blurry images.So,this paper studies spatially varying blurry images and designs an end-to-end blind restoration network based on a dual-attention dynamic filter network.The method proposed in this paper has abandoned the traditional blur kernel estimation strategy and achieved the effect of realtime restoration of spatially varying blurry images,which has great academic significance and practical value.The main work of this paper is as follows:(1)Researched the status of blurry image restoration native and foreign and summarized the research difficulties of this topic,analyzed the related theories of blurry image blind restoration and convolutional neural network,deeply analyzed the blurry image blind restoration methods,and summarized the shortcomings of the existing methods.(2)A blind restoration network of spatially varying blurry image based on dual stream dynamic filter network was proposed.In most existing networks,the convolution kernel with the same weight is used to filter in different spatial positions when restoring blurry images,which ignores the differences between different blurry pixels and cannot handle complex spatially varying blur well.To solve this problem,this paper introduced the idea of dynamic filter network,and proposed a dual stream dynamic filter network for spatially varying blind image restoration.The network is composed of a dense dynamic local filter network and dense residual generation network.Dense dynamic local filter network is used to generate the one-to-one dynamic convolution kernel for each pixel to realize the per-pixel dynamic filter operation.The convolution kernels generated by dense dynamic local filter network dynamically adapt to the input blurry images and the spatial positions of pixels,which can handle different types of spatially varying blurry images with different degrees of blur.Dense residual generation network is used to make up the receptive field of dense dynamic local filter network to a certain extent and further improve the restoration effect.Finally,the effectiveness and feasibility of the method in this chapter are verified by the relevant verification experiments on Go Pro dataset.(3)A blind restoration network of spatially varying blurry image based on dual attention dynamic filter network was proposed.The spatially varying blurry image restoration network based on dual stream dynamic filter network proposed in(2)still has many deficiencies,the difference between the pixels of different channels is not considered,and the receptive field is limited.In order to solve above problems,this paper carries out further research and proposes a blind restoration method of spatially varying blurry image based on dual attention dynamic filter network.The network consists of two sub-networks: a multi-scale dynamic filter network and dual attention enhanced residual network.The multi-scale dynamic filter network generates the one-to-one dynamic convolution kernel for each blurry pixel of different channels,and realizes the per-pixel dynamic filter operation.The restoration ability of the model is improved by combining the multi-scale restoration strategy.The dual attention enhanced residual network increases the receptive field of the model and captures the global context relationship of the blurry image by introducing the spatial self-attention mechanism and channel selfattention mechanism at the same time.In addition,the dual attention enhanced residual network uses the receptive field selection module designed in this paper to better learn the nonlinear features of spatially varying blur,which dynamically fuses different receptive field features,and effectively improves the nonlinear representation capability of the network.The output features of the two subnetworks are fused to effectively integrate the local reconstruction ability and the global information aggregation ability of the restoration network,and strengthen the dynamic restoration ability of the network to different types of spatially varying blurry images with different degrees of blur.Finally,the effectiveness and feasibility of the method in this chapter are verified by the relevant verification experiments on Go Pro?Video Deblurring?REDS datasets.Experimental results show that the dual attention dynamic filter network proposed in this paper has better restoration effect and stronger robustness.