Two-Frame Convolutional Neural Network For Blind Motion Image Deblurring

Image motion blur is mainly caused by relative motion between camera and camera,or camera shake.The blind motion image deblurring refers to estimating the original unknown image from a degraded observation when the blur kernel is unknown.It is known that it is a ill-posed inverse problem.Blur kernel estimation is one of the key problem in image blind deblurring and is also a bottleneck restricting the application of blind motion image deblurring.The Convolution Neural Network in deep learning can realize the spatial mapping of the blurred image to the restored image through the convolution kernel.At present,it has become an important development direction in the image processing research community.Therefore,a two-frame convolutional neural network is proposed for the blind motion image deblurring,which has no blur kernel estimation.The idea comes from generative adversarial nets(GAN)and it restores clear images directly by applying adversarial training between the generative model G and discriminative model D.Because the proposed network is an end to end the blind motion image deblurring method without estimating blur kernel,it can greatly improve the quality and speed of the image restoration.Therefore,the research in this paper has great academic value and practical significance.The main works of this paper are as follows :(1)The article investigated the theory of blind motion image deblurring to analysis the characteristics of motion blurr image and kernel,and summarize the basic principles and difficulties of image degradation models,noise types,further studied the basic theory of convolution neural network and blind motion image deblurring methods using deep learning.(2)To achieve the end to end blind motion image deblurring,this article improved GANs derivation models.The main procedure was to make an in-depth study of the GAN,compare the advantages and disadvantages of several GANs derivation models,and apply the improved the deep convolutional generative adversarial network(DCGAN)and Least Square Generative Adversarial Net(LSGAN)in the derived model to blind motion image deblurring.Through the experiment,it was found that DCGAN was prone to gradient saturation,and LSGAN would change the restoration image style because of the lack of image fidelity items.(3)This article proposes a two-frame convolutional neural network consisting of G and D.Among them,the G is a fully convolutional networks inspired by Vaccination Gruidelines Group(VGG)and Resnet,which has added layers and improve training speed.The D like a two classifiter network,which is simple and flexible.In order to get optimal training results,the loss function completely adopts the minimum mean square error and replaces the cross-entropy commonly used in the classifiter network.The advantage is to avoid the disappearance of the gradient and maximize the optimization efficiency.At the same time,the G's loss function add the image fidelity term to assure that the constrained restoration image more similar to the clear image.(4)Designed and verified relevant experiments.The experiment was divided into two parts: the first was the performance test of the proposed two-frame convolution neural network.The second was comparing with the present excellent traditional methods and the blind motion image deblurring methods of deep learning.The related experiments in this article showed that the proposed two-frame convolutional neural network is simple and the model is easy to solve,and it has high quality and fast speed for blind motion image deblurring.