Research On Image Deblurring Methods Based On Generative Adversarial Networks

Digital images are indispensable and ubiquitous in daily life and scientific research.With the rapid development and widespread popularization of intelligent technology and terminals with cameras.People record important moments in their work and life through images,which can carry a huge amount of information.Image deblurring is a basic and challenging task in the fields of computer vision task.However,images are inevitably blurred due to uncontrollable factors such as unconscious object movement and camera shake during exposure.The image deblurring problem is to estimate an unknown clear image from a given blurry degraded image and provide clear semantic content and detailed information for high-level computer vision tasks.In recent years,deep learning has been widely used in image deblurring,and image deblurring technology has made considerable progress.Based on this,this dissertation studies image deblurring from multiple angles for focusing on restoring the semantic content and fine details of the image.The main work of this dissertation is summarized as follows:(1)Image deblurring can be regarded as the cross-domain mapping learning between blurry images and clear images.This method proposes an image deblurring model with enhancement based on Generative Adversarial Networks.The model outputs deblurring images directly in a pure data-driven learning mode without introducing accumulated errors caused by kernel estimation and non-blind deconvolution operation.By constructing a two-stage generator including image encoding and decoding as well as enhancement module,the capability of constructing high-order residual functions and complex features of the network is promoted,and the clarity of images processed by the network is improved.(2)An image deblurring method based on perceptual features and the multi-scale network is proposed.The multi-scale generator is used to extract the features of images at different scales,so as to improve the feature expression ability of the network.In addition,the perception feature of the image is taken as a global prior,and the optimization of multiple loss functions constraint network is introduced from the dimension of image content,structural and detailed,so that generated images having a significant structure and good visual effect.(3)The existing image deblurring methods obtain the features of the image mainly through local receptive fields and model the essence of image blurring,but the nonlocal features representing the overall data distribution of the image have not been considered.In order to solve this limitation,combining with the local and non-local features of images.We consider the dependence of the image space of global and local neighborhood spatial relationships within receptive fields,this method proposes an image deblurring method based on an attention mechanism,by explicitly constructing the interdependencies among channels,pixels,and scales,this method improves the network’s ability to learn complicated features,therefore obtaining high-quality recovered images.(4)The existing deblurring methods based on Generative Adversative Networks only focus on the design of architectures and loss functions of generators,while the network can not obtain deblurred images with prominent edges and fine details only by discriminating image contents.To solve this problem,an image deblurring method based on edge adversarial mechanism and partial weight sharing network is proposed.This method takes image edge information as priors and introduces an image edge reconstruction constraint term and an image edge discrimination constraint respectively for the generator and the discriminator so that the whole network can learn image edge information more consciously.Besides,a generator architecture based on partial weight sharing is proposed,which provides clear features for image detail restoration constantly by sharing the feature decoding process of clear images and blurring images.The network can recover images with salient edges and fine details.