Research On Image Denoising Method Based On Convolutional Neural Network

Image denoising is a very classic topic in image processing and a crucial step in visual fields such as object detection and tracking.It is widely used in fields such as biology,medicine,and military industry.Noise in images is a factor in images that prevents people from receiving image information.Based on the principle of image generation and transmission process,noise inevitably exists in digital images.Therefore,the study of image denoising has both high academic value and practical significance.Image denoising is the process of removing noise from an image and restoring the original image as much as possible.In recent years,convolutional neural networks have made a great splash in the field of image processing,and image denoising based on convolutional neural networks has achieved unprecedented results.However,on the one hand,most deep learning-based image denoising methods remove image noise while losing some of the original features of the image,resulting in image quality degradation.On the other hand,convolutional neural networks rely on large amount of data for model training,which is memory-consuming and time-consuming for data processing,resulting in lower image denoising efficiency.To address these problems,convolutional neural networks are studied to improve the denoising performance and denoising efficiency of images,and alternatively connected two-branch convolutional neural networks and master-slave depth-separable convolutional neural networks are proposed for image denoising,respectively.Specifically,the alternatively connected dual-branch convolutional neural network consists of an alternatively connected branch and a wide-path feature extraction branch to form a dual-branch network,and is cascaded with a self-attention mechanism.First,the dual-branch convolutional neural network can extract more noise features than a single network structure,thus recovering a more realistic image.Second,the use of residual structure in the alternating connection module improves the network generalization performance during training and accelerates the network convergence.Finally,the combination of self-attention mechanism and convolutional neural network can focus on both global and local features of the image,which can reduce the problem of losing some information of the image during denoising and improve the denoising performance of the network.The master-slave depth-separable convolutional neural network consists of a backbone network and three branch networks,where the backbone network consists of depth-separable convolutional modules and the three branch networks are densely connected network,multi-scale feature extraction network and multi-connected network,respectively.The backbone network extracts the image noise features and then uses the residual structure to obtain the real image,while the branch network extracts the image edge features,texture and multi-scale features to complement the image features in order to obtain the closest real image.The backbone network uses a depth-separable convolutional module to ensure the image denoising effect while reducing the time for training the model and improving the efficiency of model training.The three branch networks adopt dense connection,multiple connection and null convolution to further improve the denoising ability of the network.Through in-house ablation experiments and comparison experiments with advanced methods in terms of peak signal-to-noise ratio,structural similarity and training time,the two denoising networks proposed in this paper have improved over current denoising methods in terms of performance and efficiency of denoising,respectively.The average peak signal-to-noise ratio of the alternatively connected two-branch convolutional neural network on the Set12 dataset has reached 33.34 d B when the noise level is 15.The training time of the master-slave depth-separable convolutional neural network is reduced by nearly 30 hours compared with that of the ordinary convolutional neural network.The proposed method in this paper has proved to be well advanced.