Research On Deblurring Algorithm For Dynamic Scene

Image deblurring has always been an important research area in the field of computer vision.With the recent development of deep learning technology,more and more deblurring algorithms based on deep convolutional networks have come out.These algorithms input a blurred image through a convolutional neural network to output a clear image.This low-level visual task is naturally consistent with the adversarial generation model.Also,it has been shown in many visual tasks that the feature pyramid network has proven to be able to make good use of contextual information and multi-scale information to obtain good visual effects.Therefore,this paper takes its own strengths and chooses the generative adversarial network as the framework of the overall network,and the feature pyramid network as the basic framework of the generator.As well as made two innovative designs for the generator:(1)During the decoding process of the feature pyramid network,as the middle layer information is continuously integrated,its proportion gradually increases,which leads to the phenomenon that the semantic information is diluted,so it is impossible to use the global detailed information in the data through the semantic information to achieve a high-quality deblurring effect.In order to solve this problem,part three of this paper designs a semantic supplement network which incorporates the mechanism of semantic supplement in the process of generator decoding.In addition,during the decoding process,a more effective upsampling method is incorporated to better restore the details of the blurred image.(2)Many experiments show that not every channel in the deep features of the convolutional neural network can correctly learn the semantic information,even though these channel slices will receive equal attention from the network,which leads to a phenomenon that the wrong feature information misleads the network.Therefore,in order to increase the accuracy of semantic information,the fourth part of this paper designs a semantic correction network.In the encoding process,a semantic correction mechanism is added.Unlike the previous treatment of each channel slice equally,this mechanism allows the network to give the correct feature information a larger weight,thus achieve the purpose of filtering out low-value or even wrong feature channels.In this paper,a detailed experimental analysis is performed on two public data sets.The innovative algorithms designed in the paper are compared with many existing excellent algorithms.From the experimental results,we can know that the algorithms designed in this paper compared with quantitative comparative analysis,it has significant advantages.From the experimental results,we learned that the algorithm of this paper has significant advantages compared with quantitative comparative analysis.