Deep Models For Video Image Deblurring Using Discriminative Priors

As the important carrier of information reception and transmission,images have penetrated into all fields of social life.During the imaging process,a degraded image is generated due to the deficiency of the imaging device or other external factors.Motion blur is often inevitable under long exposure interval,especially as using hand-held cameras,which is one of the most basic and most significant research subjects in image restoration.In the framework of adversarial learning,this paper aims to explore the end-to-end approach to non-uniform blind deblurring with discriminative priors.The main works are provided as following:First of all,based on the idea that a good blind motion deblurring prior model should prefer clear images to the blurry ones,this paper attempts to incorporate discriminative priors into the neural network in the form of loss functions,so as to further improves the deblurring performance by combining model-based and data-driven strategies.Specifically,this paper studies two such loss functions,based on a discriminative gradient prior(DGP)and an opposite-channel-based energy(OCE),respectively.Without any adjustment on the network architecture of Deblur GAN,experiments show that OCE can effectively enhance the deblurring performance.Secondly,during the training of Deblur GAN,it is discovered that Deblur GAN is inherently problematic in its generator,naturally restricting its real blind deblurring capability.Therefore,two aspects of endeavors are made for a more effective and robust adversarial learning approach to dynamic scence deblurring.On the one hand,a simple,stable and architecturally robust deep auto-encoder is developed as a substitute of the original generator in Deblur GAN,and some important components such as skip and dense connections are introduced into auto-encoder.On the other hand,in order to further improve the deblurring performance,the opposite-channel-based energy loss function proposed in this paper is plugged into the deblurring framework.For brevity,the proposed model is named as Deblur GAN+ in this paper.Deblur GAN+ is proved to be superior to Deblur GAN on both large-scale dynamic scene blur datasets and real blurry scenes.Finally,this paper makes a critical reflection on the learning-based blind deblurring methods,mainly from two aspects of datasets and models.On the one hand,in the era of deep learning the dataset is an important indicator for evaluation of results,and hence this paper firstly makes a critical analysis on the rationality of the classic dataset construction process.On the other hand,this paper futher evaluates the practicability and robustness of recent representative deep deblurring models.The dataset used for testing not only covers the real dynamic blurred scenes,but also the synthetic and real static blurred images,in order to comprehensively evaluates deblurring performance of various deep deblurring methods in different types of blurring scenes.The results show that synthetic blurry images cannot replace real blurry images during learning of deep deblurring models in the current,though the synthetic methods of blur datasets are becoming more and more reasonable.Meanwhile,it is found that the performance of state-of-the-art deep deblurring models is even far behind those traditional model-based methods on the static blurry images.Hence,there is still a large research space for deep deblurring methods.