Research On Image Deblurring Algorithms Based On Deep Learning

Image is an important carrier for recording and transmitting information in daily life.In the process of acquisition,storage,and transmission,the image will suffer from some problems such as deterioration of quality,loss of information and other issues.We call this phenomenon image degradation.Among them,image bluring is one of the main types of image degradation.The causes of image bluring can be divided into defocus bluring(caused by the camera's inaccurate focus)and motion bluring(caused by changes in the relative position of the camera and the imaging target,such as displacement,rotation and so on).Image deblurring is an important branch in the field of computer vision.Its purpose is to restore blurred images to clear images and provide effective information for subsequent image analysis and understanding.Therefore,the research of image deblurring technology is of great significance.In recent years,deep learning has been widely applied in the field of image deblurring.This thesis will carry carry out image deblurring research based on deeplearning.The main research work is as follows.(1)A single image deblurring algorithm based on Generative Adversarial Network is studied.In this thesis,Cycle GAN is used to perform end-to-end blind deblurring of unpaired data,that is,a single image,and an appropriate loss function is proposed to constrain the generated clear image.Because the feature information of a single image is limited,in order to make full use of the information,the generator is transformed into multi-scale recursive network.At the same time,dense blocks are introduced to replace residual blocks in the multi-scale recursive network to achieve the sharing of parameters during scale iteration,reducing the amount of parameters effectively,lowering the difficulty of training,and improving the stability of the network model.Experimental results show that the network is superior to other algorithms in single image deblurring,and is also better than the existing Deblur GAN,DMPNH and other methods when compared with image deblurring algorithms in paired data sets.(2)A video deblurring algorithm based on multi-scale recursive network is studied.A video deblurring can be regarded as multi-frame images deblurring.This thesis uses a multi-scale recursive network to construct a video deblurring network,integrates octave convolution simultaneously,decomposes the video frame into high-frequency and low-frequency parts,and processes the features of high-frequency and low-frequency parts separately to ensure that under the full use of feature information,In the low-frequency part processing,the amount of calculation is reduced,resources are saved,and the redundancy of the network is reduced in the low-frequency part processing.Experimental results show that this algorithm improves the video deblurring effect and increases the training speed.(3)An image deblurring system is designed and implemented.The system is based on the architecture of client side and server side implemented by Web.The client uploads a blurred image,and the server returns a clear image to the client after deblurring.This project realizes the research of image deblurring algorithm.