Research On Deblurring Algorithm Of Motion Video Image

There are many ways of transmitting information,such as language,text,image,video and so on.With the development of intelligent devices,it is more convenient to access these information carriers for people.Image plays an important role in the process of information transmission by the characteristics of carrying large amount of information,easy to obtain and so on.However,the camera equipment would get blur image due to movement during the process of acquiring the image and make the image lose some useful information.This is where image deblurring comes in and it is widely used in traffic monitoring,medical imaging,photography and other fields.The traditional image deblurring algorithm is not effective because it is difficult to estimate and produce the blur kernel accurately.As scholars apply deep learning technology to image deblurring research,the field of image deblurring has gained considerable development and image deblurring algorithm has excellent deblurring effect.The main work contents of this thesis are as follows:(1)Based on deep learning,combined with generative adversarial learning,residual learning and multi-scale feature fusion theory,a multi-scale and patch-stacked image deblurring neural network was proposed.The input data of each scale of the network is reduced by direct segmentation which avoids huge computation amount brought by the conventional subsampling method.A residual block is designed and embedded into the shallow part of the convolutional neural network,which is combined with the dual attention mechanism,so that the network can pay attention to the useful information of the input image at different scales,and enhance the representation ability of the model.Experiments show that this method can effectively improve the deblurring performance and efficiency of the network.(2)In this thesis,two image deblurring methods combining time information are proposed.One is the image deblurring algorithm in series of the traditional optical flow estimation network,and the L-K optical flow estimation method is used to preprocess the input image and then send it into the pre-trained deep learning-based image deblurring network.The other method is to cascade Liteflownet optical flow estimation network,which uses the same pixel points on the front and rear frames as well as the intermediate frames to assist the training process to obtain the change information of the same pixel in continuous time in order to improve the reconstruction quality of the deblurring image.We use generated adversarial network for training,and use Charbonnier loss function and edge loss function to solve the problem that the subjective effect of image reconstruction is too smooth due to the commonly used L2 loss function.Experimental results show that the algorithm can further improve the quality of image reconstruction.