Research On Motion Blurred Image Restoration Algorithm Based On Deep Learning

Motion Blurred image restoration is the process of restoring the blurred image to a clear image by using the relevant knowledge theory of the image,mathematical principles and computer technology to reduce image noise,blurring,etc.It is also an important research direction in computer vision,image processing and other fields.The traditional motion blurred image restoration is to restore the blurred image through deconvolution to obtain a clear image on the premise of a clear blur kernel,but the method of estimating the blur kernel is not practical.In recent years,there are many methods of using deep learning to restore motion blurred images,but there are problems,such as poor recovery of details and low generalization performance.The advantage of using the demotion-fuzzy method based on deep learning in this paper is that it can improve the details of image restoration and model generalization performance:(1)A motion blurred image restoration algorithm based on multi-scale recurrent network is proposed.A multi-scale feature fusion module is introduced in the encoding terminal to increase the receptive field and obtain more feature information that is not easy to be found in the image.A multi-scale residual dense connection module is proposed,which can effectively combine local information and global information the image to ensure the integrity of information during model training.The channel and spatial attention mechanism modules are introduced,and the multiscale residual dense connection module integrated with the attention mechanism is used as the basic structure of the network to accurately collect the information of objects.Through experimental comparison,the proposed algorithm compared with other algorithms,it has higher image quality,better performance in two different evaluation indexes,and better generalization performance.(2)A motion blur image restoration algorithm based on generative adversarial network is proposed.The feature fusion module of the generated network is improved,a new residual module is proposed,and activation function is used to solve more nonlinear problems and learn more complex feature information.The deep separable convolutional neural network is introduced to reduce the number of parameters and improve the training speed.Moreover,the deep separable convolutional neural network is combined with the new residual network,and the deep separable residual network is proposed as the main part of the feature fusion module,which can improve the network performance while reducing the amount of parameters.In the experimental testing phase,three different datasets were selected for testing,and compared with other algorithms,both the restoration of image details and texture and the evaluation indicators were better reflected.