Research On New Blind Deblurring Methods For Textual Images

Image is one of the indispensable carriers of receiving and transmitting information in human daily life.However,in the process of imaging,it's possible to find the decline of image quality due to several factors.The image deblurring techniques studied in this thesis is to recover clear images from blurred images.At present,the techniques are widely used in military,security monitoring,transportation,astronomy and other fields.In this thesis,under the theoretical framework of maximum posterior probability estimation and deep neural network,two new blind deblurring algorithms are proposed when textual blurred images are taken as the research object.The main results are provided as follow:1.Aiming at the special priors of the textual images found in the experiment,this thesis proposes a new image blind deblurring algorithm named L0-RTV,which combines the edge-aware image gradient L0 norm sparse model with the relative total variation(RTV)model.L0-RTV uses the L0 norm to extract the salient edges of the image,both with the RTV to distinguish the edge structure from the faint texture so that can prevent the loss of the weak edge,improving the accuracy of the kernel estimation,and making the deblurred image more realistic and natural.In order to evaluate the performance of L0-RTV,the model is compared with two most representative blind deblurring algorithms on the synthetic and real textual images.The experimental results show that L0-RTV can estimate exact kernel and recover more natural images with strong robustness.2.A new deep model named Deblur-ResNext is proposed in the thesis when using the existed deep models to deal with the blurred textual images.The model has two major characteristics.The first is to introduce an improved residual network ResNext module in the convolutional layer,and the second is to optimize the residual module structure and network parameters according to the specific requirements of the experiments.Through these,the model training speed is improved while reducing the amount of parameters.On both the synthetic and real blurred images,the comparison between the proposed model and the current two representative neural network models proves that the Deblur-ResNext network model has excellent deblurring results and fast speed of convergence.3.Most existed deep methods for blind image deblurring learn the law of images' distribution during training and estimate the blurred images during testing.Therefore,the size and distribution of the sample in training set will make a big difference for the models.The experimental objects inthis paper are license plate images and text images.It is difficult to obtain corresponding clear images and blurred images at the same time,so this paper realizes data augmentation and improves the estimation accuracy of the model by synthesizing blurred images.According to image degradation model,the synthesis of blurred images can be implemented in three steps: taking a clear image,randomly generating a blur kernel and convoluting to generate a blurred image.Part of the license plate data set used for training during the experiment was derived from the existing database,and the other part was synthesized.