Non-Blind Image Deblurring Method By The Total Variation Deep Network

The popularization of smart portable devices has dramatically changed the way people communicate.From written texts to other mediums of communication such as photos and short videos,the ongoing evolution allows emotions to be expressed in a more accurate and comprehensive manner.However,an image would lose part of its information if it is irreversibly blurred resulted from its long history or shooting jitter.Accordingly,the image-deblurring technology has become more important in terms of image application.This paper conducted in-depth research on image-deblurring technology with emphasis on:(1)Instead of sticking to conventional image deblurring based on the regularization theory model,this paper proposed a brand-new model built upon the total variation deep network.Many non-blind image-deblurring methods have already existed nowadays,especially the ones based on Total Variation(TV)models.However,how to adaptively select parameters and improve regularization becomes a critically open-ended question.The model proposed in this article can adaptively update the parameters during iterations to obtain the optimal results.Specifically speaking,this model mainly applied deep learning and prior knowledge to build a deep network on the basis of total variation,and calculated regularized parameters such as bias and weight.By incorporating the idea of deep learning to complete the automatic update of parameters,the model could avoid complex hand calculations and dramatically reduce the participation in manual debugging.Compared with several other classical methods,the model generated better results in retaining details and anti-noise performance.At the same time,the model requires relatively less training sets to achieve the same effect as that produced by a large number of training sets,thereby speeding up the calculation.(2)For those devices with limited computing capabilities,such as embedded devices and mobile devices,an efficient convolutional neural network architecture is one of the most critical and rigid requirements.In recent years,although very deep convolutional neural networks which are specific to improving image restoration have been proposed and achieved good results,their hardware requirements are often difficult to meet,thus leading to inefficiency in practical performance.This paper proposed the use of residual dense blocks,a method to ensure in-depth supervision of convolutional neural networks,sufficient gradient flow,and feature reuse capabilities.By lowering the number of parameters and calculation operations,the approach reduces costs of training anddeduction process remarkably as well as achieve effective image recovery without using any specialized software or hardware equipment.Experimental results showed that the proposed architecture is capable with respect to model size,required parameters,and even image restoration.