Research On Image Deblurring Algorithms Based On Generative Adversarial Networks

With the continuous development of deep learning algorithms in the field of image deblurring,the main idea of researchers is to improve the effectiveness of the algorithm deblurring by deepening the network,however,this idea makes the complexity of the network algorithm high,the demand for arithmetic power is relatively large,and also makes the deblurring work consume a lot of time,neglecting the speed of the algorithm,and the practicality is poor.Therefore,in this paper,we use generative adversarial networks to study around the image deblurring problem,and the main work includes.(1)Generative adversarial network image deblurring algorithm based on bidirectional feature pyramids.In order to improve the effect of image deblurring and enhance the detail information of the generated image.Firstly,a bi-directional feature fusion feature pyramid network is introduced in the generator part.Based on the traditional feature pyramid network,a bottom-up channel is added in this paper,which can improve the utilization of the underlying features and enhance the detailed texture information,and the channel can improve the bottom-up information transmission efficiency,which can make the training process converge faster and better.Then,an improved Res Next network is proposed as the backbone network of the model in this chapter,and an effective channel attention mechanism network is introduced.This channel attention mechanism obtains the importance of each feature channel during processing,and sets the weights according to the importance and relevance of the channel to optimize the extraction effect.Finally,the discriminator replaces the Batch Normalization(BN)with Group Normalization(GN)based on the use of Patch GAN with the Leaky Relu activation function.After completing the above improvements,this paper has conducted comparison experiments with several more common image deblurring algorithms and achieved better experimental results for image restoration quality.(2)Depth-separable convolution-based image deblurring algorithm for generative adversarial networks.In order to better solve the problems of slow speed,large computation and many parameters,which are common in conventional image deblurring algorithms,this paper proposes a generative adversarial network based on depth-separable convolution for image deblurring work.In Chapter 4,based on the improvement of Chapter 3,the backbone network is designed as a depth-separable convolution-based network.First,a layer of 1×1 convolution is added before the depth convolution to expand the feature channels,then,the Relu activation function in the last layer is removed and replaced with a linear function output,and finally,the Relu function in the middle is replaced with Relu6.The purpose of this improvement is to reduce the information loss and improve the detail quality of the generated images to achieve better image deblurring results.In the final experimental comparison,the improved method proposed in this chapter shows a very obvious improvement in model size and computational speed.