Research On Single Image Blind Deblurring Reconstruction

Image deblurring is always the focus of image processing and computer vision.Traditional deblurring methods based on statistical priors depend on the selection of priori types to a great extent,whichs detail recovery ability is limited.At the same time,the deblurring model is difficult to solve and its operation is complex.In recent years,deep learning has become a research hotspot in various fields.Because convolution neural network(CNN)can fully mine the internal features and prior knowledge of the image,this paper studies the image de-blurring method based on CNN.The specific work is as follows:In order to solve the ill-posed problem of image restoration,a de-blurring model is usually constructed by using prior information as a regular term.The traditional methods based on statistical priors are sensitive to noise and have limited ability to restore details.Aiming at these problems,this paper deeply studies the key technologies of denoising CNN,designs a fast denoising convolution neural network,and learns the image denoising priori.Finally,an image blind deblurring model is constructed based on gradient sparse prior,depth denoising prior and low-value pixel prior,and an optimal numerical algorithm is given.Experimental results show that the proposed algorithm not only has good detail recovery ability,but also is more robust to the image and its blur type,noise level and so on.Compared with the existing mainstream algorithms,this method has obvious advantages.End-to-end convolution neural network can learn the pixel-level mapping relationship between degraded image and clear image,and can effectively remove spatial variable blur.However,the conventional convolutional neural network has the shortcomings of insufficient generalization ability and loss of detail information,which leads to over-flat restored image and insufficient detail recovery.In this paper,a parallel network structure is used to decompose the image deblurring task into two subnetworks: deblurring and detail recovery.At the same time,a multi-scale network input structure is designed to obtain the blurred image information to increase the network receptive field and obtain the detail features of different levels of images.Then an image de-blurring method based on multi-scale and multi-task CNN is proposed.The experimental results show that the details of the deblurring image obtained by the network are richer,which indicates that it is suitable for removing spatial variable blur and its deblurring performance is superior.