Research Of Key Techniques For Image Deblurring Based On Prior Learning

Recently,images have become an important way for people to obtain information and communicate with the rapid development of modern society,especially the popularity of smart phones and other portable mobile imaging devices.However,the process of imaging is never perfect,and there are some uncertainties in the measurement process in the form of blur,noise,and other degradation factors.Image deblurring is to restore potentially shape images from blurred observations,which is a very challenging problem in image processing field.Modeling the prior of shape images is the core of the image deblurring problem,although many effective prior models has been presented,there is still not a perfect prior which can capture the rich semantic information of natural images,so combining with the development of current technology(especially the rapid development of deep learning),it is important to further explore the prior model.In this paper,we adopt a research strategy of image priors from non-blind image deblurring to blind image deblurring.We firstly explore the image priors based on the statistical characteristics of shape images,gradually explore the deep priors based on the data driven,and finally search for the key components affecting the performance of image reconstruction(named the structure prior).The main works can be described as follows:(1)For the prior based on image statistical features,a novel efficient method is proposed to promote hyper-Laplacian prior for the deconvolution of blurry and Poisson noisy images.We firstly analyze the mathematical characteristics of generalized lp/lq norm based on the Berkeley Segmentation Data Set(BSDS)and implement deconvolution using generalized lp/lq norm for derivatives in high frequency domain,so as to enforce the preservation of strong edges.Then hyper-Laplacian prior is promoted for actual Poisson image deconvolution,which uses the estimated gradient information to penalize the small derivatives(often noise)and preserve large derivatives associated to image borders.Our preliminary experiments show that proposed method can achieve higher image quality than state-of-the-art methods in terms of Peak Signal-to-Noise Ratio(PSNR)and noise sensitivity.(2)Combining with the mathematical characteristics of generalized lp/lq norm,its validity applied to the image blind deblurring problem is explored,and then an efficient method is presented for blind image deconvolution.Firstly we analyze the mathematical characteristics of generalized lp/lq norm,then apply generalized lp/lq norm-based prior model in gradient space to estimate the blur kernel.Due to the complexity of optimization model,we develop an alternating gradient descent method to solve the generalized lp/lq norm-based model which can achieve high recovery quality.Specifically,the selection strategy of regularization parameters is given by using generalized cross-validation method,and these parameters can be updated in alternating minimization steps.Our preliminary experiments show that the proposed method can achieve state-of-the-art results.(3)For exploring the correlation between traditional image deblurring methods based on energy optimization and deep learning,we propose a neural generative network for object motion deblurring by using deep image prior(DIP).To this end,a local blur detection-based neural network is proposed,which is composed of three generative networks to model the deep priors of clean image,blur kernel and the weight variables of object motion images respectively.Our preliminary experiments have been conducted for static scenes and object motion scenes deblurring,which show that the proposed method can achieve notable quantitative gains as well as more visually plausible deblurring results compared to state-of-the-art methods.(4)With the effectiveness of deep generative prior(DGP),we present a general solution for single image deblurring using DGP as the image prior.To this end,two aspects of this object are investaged.We firstly model the process of latent image degradation,corresponding to the estimate of blur kernels in conventional deblurring methods.In this regard,a Reblur2Deblur network is proposed and trained on the large-scale datasets.Then we implement the proposed deblurring framework with a relaxation strategy for tackling this difficulty.The pre-trained GAN and ReblurNet are allowed to be fine-tuned on-the-fly in a self-supervised manner.Finally,we demonstrate empirically that the proposed model can perform favorably against the state-of-the-art methods on benchmark datasets and real-world blurry images.(5)To explore the structural prior of the latest end-to-end deep deblurring network,we investigate two aspects of network architecture design for dynamic scene deblurring.We firstly learn blur characteristics and their location in dynamic scenes,which corresponds to learning what and where to attend in the channel and spatial axes,respectively.In this regard,we design an attention-adaptive module(AAM),the innovation of which is that it adaptively determines the arrangement of channel and spatial attention modules(i.e.,sequentially or in parallel).Then we propose a deformable convolutional module(DCM)to handle geometric variations.Preliminary experiments demonstrated that incorporating the AAM and DCM into existing deblurring models can significantly improve performance.