| Image is ubiquitous in human’s daily study and life,and is the most commonly used carrier to spread information.Due to the influence of shooting environment,imaging equipment and other factors,image noise is inevitably introduced in the process of imaging and transmission,and image noise reduction is a classic topic in the field of computer vision.How to effectively remove noise and maintain image structure information and detailed features is not only the premise of criminal investigation and case solving,satellite remote sensing image,medical image processing and other fields,but also the key to the smooth progress of image recognition,image classification,image segmentation and other research work.With the rapid development of deep learning technology in the field of image processing,convolutional neural networks have achieved higher and higher performance indexes for artificial added noise and real noise.However,how to design a network model to simultaneously meet the needs of two types of noise,how to reduce the loss of detailed features in the training process,how to design loss function to make the training process efficient,There is still room for improvement to reduce image edge blur.In view of the above problems,this paper mainly does the following work:1.The research background,significance and development status of image noise reduction technology are briefly summarized.The existing image noise reduction algorithms are classified and summarized,and the performance of several common methods are compared.Then it introduces the basic knowledge of deep learning network in image denoising and the application and development of convolutional neural network in image denoising network.2.In view of the problems that traditional deep residual networks cannot effectively solve the complex real noise of the model,and the model generalization performance and model robustness of the real noise need to be improved,an improved multi-scale deep residual network model is proposed.The feature extraction task of the image is divided into three parts: encoding,decoding and skipping.Firstly,the convolution kernel of three different scales is set for downsampling to complete the multi-scale feature aggregation.At the same time,the channel attention mechanism is added to screen out useful features and pass them into the next scale..In the decoding part,feature dimension is raised through upsampling,and features of the same scale are superimposed and supplemented in the encoding stage and decoding stage through cross-connection.Finally,the input image and noise features are calculated through the learning mapping of residual unit to obtain the de-noised image.3.A multi-scale depth residual noise reduction model with multi-stage supervision is proposed to solve the problem that the existing multi-scale convolutional neural network noise reduction model ignores the feature extraction of the original resolution image in the process of convolution,which leads to the loss of details.On the basis of Work 2,the input image is input in stages,and the input depth residual network model of different sizes of image blocks is used to expand the perceptive field.The guiding features are screened out by the supervisory attention mechanism and passed into the next stage to improve the overall denoising performance of the model.4.In view of the traditional mean square error loss can only be compared and determined at the pixel level,which has excellent performance in the artificially added noise.In the real image noise task,more advanced semantic feature level learning cannot be carried out in the task of real image noise,a joint loss function is proposed,which can remove different types of noise by setting super parameters.It not only generates clearer target images,but also improves the applicability of the model.In the experiment,BSD400 data set was selected for training,and Gaussian noise reduction test was carried out by Set12 data set.The real noise reduction test was completed by SIDD data set.The comparison with common noise reduction neural networks shows that when Gaussian noise σ =15,25,50 is added to the image,the PSNR of the image after noise reduction by the proposed algorithm is improved by 0.03 dB,0.05 dB,0.14 dB compared with DNCNN with good performance of Gaussian noise elimination.At σ= 25,50,the results are 0.02 dB and 0.06 dB higher than those of MPRNET,respectively.For images with real noise,the PSNR of the image denoised by the proposed algorithm is 0.48 dB higher than that of the CBDNET algorithm.Experimental analysis shows that the proposed algorithm has high robustness in image noise reduction,which can not only effectively recover image details from noise,but also fully maintain the global dependence of image. |
