Research On Image De-motion Blur Algorithm Based On Deep Learning

In the imaging process,due to the equipment,weather,motion,noise,etc.,the obtained images are degraded,blurred,or even unable to recognize effective information,which cannot meet the needs of subsequent applications.Therefore,it is extremely important to process these blurred images.Motion blur is a type of blur caused by the relative movement between the imaging device and the shooting target.The research interest in this thesis lies in the processing of motion blurred images.In the thesis,firstly,the mathematical model of blurred image restoration is introduced.Then,the common types of blur and de-motion blur methods are researched,as well as analyzing their advantages and disadvantages.Finally,two improved image demotion blur algorithms are presented.1.Proposed an image de-motion blur algorithm based on U-Net structure.Based on the traditional U-Net structure,the deblurring effect is improved by increasing the number of hop connections,and the improved U-Net is used as the generator;the PatchGAN network is selected as the discriminator adopts,and the number of network layers of the discriminator is modified.The presented algorithm is verified through simulative experiments,and the experimental results are compared with those of several other algorithms.Experiments indicate that the new model has better performance than other algorithms in structural similarity index and peak signal-to-noise ratio when the degree of image motion-blur is low.2.Presented an image de-motion blur method based on generation adversarial network(GAN).The new algorithm has been tried and improved on the network structure and loss function.The residual connections are introduced in the generator structure,which is pre-trained by using the dataset composed of motion blurred images and their corresponding clear images.The perceptual loss is introduced in the generator loss function,and the gradient penalty term is introduced in the discriminator loss function.The simulative experiments with different degrees of motion blur images are carried out.Experimental results show that,under larger motion blur condition,the new algorithm has better performance,and the restored image can retain more image texture details.