Research On Image Deblurring Based On Generative Adversarial Network

Image deblurring is a popular research direction in the field of image processing.The motion blur of the image will lead to the degradation of image quality and the loss of detailed information,which will have adverse effects in the fields of medical treatment,transportation,aviation,and industry.In practical applications image motion blur is often inhomogeneous and the blur scale is diverse.In deep learning deblurring tasks,in order to deal with these uneven blurs and achieve better results,it is necessary to use larger convolutional kernel size to increase the perceptual field or to increase the model depth of the network.However,the makes the deblurring model more complex,requires more computing power of the computer during training,and the efficiency of the deblurring task is greatly reduced.Existing deblurring models often suffer from severe loss of image texture edge information and poor model generalization ability.Therefore,this paper uses generative adversarial networks to analyze and study the problems surrounding deblurring.The project proposes an edge spatial attention residual generative adversarial network for image deblurring.The generator draws on the U-Net structure using an encoding-decoding structure,and introduces a spatial attention mechanism in the generator.Edge information is introduced on top of the traditional spatial attention.Edge information is extracted using a bi-directional gradient fusion algorithm for fuzzy map-clear map in GOPRO dataset and REDs dataset,and then the feasibility of introducing edge information is analyzed and illustrated.The use of edge information guides the model training,allowing the model to focus more on the extraction of edge information,which is beneficial to the model’s recovery of edge information from blurred images.The residual module using edge space attention and residual structure is used to learn the residuals of the image.In the decoding module,Up Samle layer and convolution layer are used instead of deconvolution layer to avoid tessellation effect.A jump connection layer is introduced between the encoding and decoding modules to efficiently utilize the image features extracted at higher levels.The discriminator uses the structure of WGAN and replaces batch normalization with group normalization.An edge loss function is added to the loss function,which facilitates the network model to recover the image edge information better.In the overall test of the GOPRO dataset,PSNR,SSIM,and VIF are 30.64,0.9061,and 0.5928,respectively,ranking second among the six comparison methods.The experimental analysis shows that the proposed model is at a leading level in terms of deblurring effect.The project proposes an high-frequency attention residual generative adversarial network for image deblurring.The high-frequency attention residual generation adversarial network is proposed by further improving on the edge space attention generation adversarial network.The frequency domain information and attention mechanism are combined to form a Fourier channel attention module.The fast Fourier transform algorithm is used to extract frequency domain information from the fuzzy map-clear map in GOPRO dataset and REDs dataset,and the feasibility of introducing frequency domain information is analyzed and illustrated.The frequency domain information can make the model focus more on the extraction of texture information and facilitate the recovery of texture information of blurred images by the model.The Fourier channel attention,edge space attention and bottleneck residual structure are used to form the high frequency information residual module.A structure similar to Inception is designed to replace the first feature extraction 7×7 volume data layer,which has a smaller number of parameters and a larger perceptual field.In the overall test of the text data set,PSNR,SSIM,and VIF are 28.78,0.9795,and 0.5753,respectively,and have the best restoration effect among the six comparison methods.At the same time,the proposed model improves the speed of deblurring by 9.5% compared with the edge space attention residual generation adversarial network.The experimental analysis concludes that the speed of the proposed model is better than the edge space attention residual generation adversarial network,and the deblurring effect is at a leading level.In particular,the application in text image deblurring is remarkable and exceeds expectations.