Research On The Technology And Application Of Image Blind De-blurring Based On Conditional Generative Adversarial Networks

Image de-blurring is a hot topic in the field of computer vision.In daily life,the blur image blur kernel is almost unknown,the key to solve the problem is to make full use of the prior knowledge of image.Most existing blind de-blurring methods are based on regularization or hypothesis to estimate unknown blur kernel.However,this can lead to high computing costs,checkerboard artifacts,lack of good model generalization,loss focuses on restoring texture details.Based on the Conditional Generative Adversarial Networks,a blind de-blurring method is proposed to solve the these problems and it is implemented on the mobile terminal.The main work is as follows:(1)For the existing problems of image blind de-blurring model,this thesis uses convolutional neural network as the basic architecture of generator and discriminator.The two networks against each other to solve a lot of computing costs.Using the up-sampling and convolutional layers to avoid the checkerboard artifacts of the generated image,using instance normalization to find the differences between the two image pixels,can largely to restore texture details of the blurred image.The whole model uses the Go-Pro dataset.The experimental results show that it is superior to previous work in terms of evaluation metrics peak signal-to-noise ratio,structural similarity measurement and runtime,and can restore street scene images.(2)For the problem of model transfer,this thesis freezes the pre-training model to the mobile terminal can solve the problem of wasted traffic and network connection caused by the client-server model.The final result shows that the blurred image taken by the mobile terminal can be effectively handled.