Research On Image Denoising Algorithm Based On Block Matching And Multi-scale Feature Networ

Digital images are subject to various types of noise and external factors during their creation,transmission,and storage,which directly affect the clarity and accuracy of the images.Therefore,image denoising has become one of the current hotspots of research.To address image noise pollution,researchers have mainly focused on two directions: traditional filtering and deep learning.Traditional filtering methods focus on noise filtering,but their results are often too uniform and lack high-frequency details.With the rapid advancement of convolutional neural network technology,noise separation algorithms have also received increasing attention and have quickly become a hot topic in today’s academic world.However,these algorithms still suffer from problems such as feature loss and edge blurring during feature extraction and fusion processes.To address these issues,this paper mainly conducted the following research:(1)To address the issues of the lack of analysis of the overall structure of images and insufficient preservation of texture details in traditional image denoising processes,a block matching and NLPM diffusion model-based image denoising model is proposed.The denoising process of this algorithm includes two steps: preliminary denoising and final denoising.The preliminary denoising uses block matching and 3D filtering methods to process the image and obtain the preliminary approximation result of the original image.Based on this,the final denoising process is carried out.This processing method uses the NSST non-subsampled shearlet transform to extract high-frequency sub-band information from the basic estimation image.Then,the NLPM diffusion filtering model is used to diffuse filter the high-frequency sub-bands to reduce the step and speckle effects brought by PM diffusion filtering,while effectively preserving the details and texture information in the image.Finally,the inverse NSST transform is applied to the low-frequency coefficients and the high-frequency coefficients after filtering to obtain the final approximation image result and achieve the final denoising processing of the original image.Experimental results show that the proposed model outperforms traditional denoising methods in terms of denoising effectiveness and preservation of image texture details.(2)Aiming at the problem of convolutional neural networks(CNN)unable to achieve a complex balance between spatial details and advanced contextual information in multi-scale feature extraction and fusion,leading to edge blurring,we propose a multi-scale feature network based on VGG16 for image denoising,called MSSNet.This model consists of three parallel sub-networks with shared weights.First,different scaled versions of the same corrupted image are input into the VGG16 network for feature extraction to improve the ability to obtain global features,obtaining feature maps outputted by different convolutional blocks.Then,a structured feature fusion module(SFFM)is used to refine and fuse multi-scale features from different convolutional blocks,based on conditional random fields and message passing mechanism,supplementing shallow features with deep features.Next,a channel attention weighting technique is used to perform multi-scale channel attention weighting on the feature maps.Finally,a multi-scale discriminant feature loss function(MDF)is used to calculate the feature differences between the output of each layer at different scales,and the network learns the optimal method for upsampling using transpose convolution to generate a clearer,more detailed denoised image,optimizing image perception quality and outputting more spatial information and texture details.Experimental results show that the proposed network model has achieved improved preservation of spatial detail and context information compared to other comparative models.