Image Motion Deblurring Based On Conditional Generative Adversarial Networks

Digital cameras,smart phones,etc.have been widely used in lots of aspects of people's work and life along with the development of the information technology.The images generated by these equipments provide plentiful information for human life.People usually use images to record life,communicate and exchange information.However,in our life,it is inevitable to receive blurred images due to certain factors,such as inaccurate focusing,or motion blur caused by the relative movement between the vehicle and the roadside objects during vehicle driving,which bring inconvenience to life and research.Therefore,image deblurring plays an extremely important role in traffic management,target detection,and video surveillance,etc.Traditional image deblurring algorithms often need to estimate blur kernels based on a priori assumption,resulting in high computing cost and poor model generalization ability.With the rapid development of deep learning,the performance of generative adversarial networks in the field of image processing has been continuously improved,especially the application of conditional generative adversarial networks for image deblurring has become a hot research.In this thesis,the following two methods based on the conditional generative adversarial networks are studied for the image deblurring as follows:(1)An end-to-end image motion deblurring method based on DeblurGAN was proposed with conditional generative adversarial network.The low-rank decomposition technique and global symmetric skip connection are used to accelerate the convergence of the network,and the gradient image L1 loss and the mutual information loss are used to make the edges of the generated image clearer and to retain as much original image information as possible.Finally,it was verified by experiments that the proposed method can reduce computation cost while maintaining a good deblurring effect.(2)Since the image pairs of corresponding sharp and blurred images required by the previous deblurring method rarely exist in real life,this thesis further studies an unsupervised image deblurring method based on disentangled representation.Network training does not require paired sharp-blurred image pairs,but uses unpaired image datasets.The content and blur features from blurred images are disentangled by using a content encoder and a blur encoder,and deep learning technologies such as KL divergence loss,adversarial loss,and perceptual loss are also used to optimize the model.In addition,a classifier is designed to guide the generator to learn the interpretable attributes of the input image based on the analog inference,so as to generate a sharper and more realistic deblurred image.In this thesis,supervised image deblurring based on DeblurGAN and the low-rank decomposition and unsupervised image deblurring based on disentangled representation are studied respectively.The former effectively compresses and accelerates the deblurring network under the premise of ensuring the quality of the deblurred image;the latter uses unpaired sharp-blurred image datasets for unsupervised image deblurring,which alleviates the difficulty of obtaining the image pairs of corresponding sharp and blurred images to a certain extent.