Deep Fully Convolution Neural Network Based Method For Single Image Non-uniform Blind Deblur

It is easy to obtain degraded image with complex non-uniform blur when capturing objects under poor exposure condition.In most cases,knowing the blur information in advance is hardly possible,let alone re-shooting of clear images.Therefore,the technique of single image non-uniform blind deblur has significant study value.Traditional image blur removal methods are limited in application range and difficult to meet the needs of real-time applications.Therefore,researchers have shifted their attention to deep learning based solution.However,only a few deep learning methods are designed for both non-uniform blur removal and blind deblur.Moreover,these methods still have the following drawbacks: 1)They still rely on traditional deblur methods,thus being time-consuming.2)Special design for specific blur types is needed,application range is limited.3)They require patch based methods to handle non-uniform blur,while the operation of clipping and merging image pieces increase the complexity of algorithm.To resolve the above issues,this dissertation proposed a fully convolutional network for non-uniform blur estimation and removal.The main contributions of this article are:1)We proposed a fully convolution network architecture to estimate blur parameters,called P-net.P-net is able to output pixel-wise parameters of multiple blur types under single network architecture directly,without divide image into pieces.Thus it expands the scope of parameter estimation and refines the parameter estimation granularity.2)We proposed a fully convolutional condition generation network structure to recover sharp image,called G-net.The output of P-net is used as a condition,which guides G-net to learn the operation of non-uniform blur removal.P-net and G-net are ultimately integrated into an end-to-end network called PG-net,which guarantee the consistency of parameter estimation and blur removal,thereby improving algorithm efficiency.3)We proposed two methods for generating synthetic image with non-uniform blur and construct the training dataset.The first method is based on the semantic segmentation of images,and the second one is based on multi-scale image patches.We conduct several experiments to evaluate the effectiveness and efficiency of our method.Experiment results show that our blur parameter estimate method as well as our deblur method outperforms the comparison method both quantitatively and qualitatively.