Image Denoising Based On Deep Convolution Neural Network Method Research And Application

Image denoising is the process of reducing the noise in digital image,and it is the important premise of image segmentation,edge detection,feature extraction and so on.Deep convolution neural network has achieved great success in image recognition,speech processing and other aspects.Through deep learning,the machine can simulate human's vision,hearing,thinking and other behaviors,and can overcome the difficulty in dealing with complex problems in pattern recognition.Image denoising has a great application in video analysis and speech processing.Deep convolution network makes use of the advantages of deep learning without statistical method to analyze data,and has achieved success in image denoising.In view of the shortcomings of traditional image denoising methods based on block matching,which can only deal with 2D image and has low denoising performance,an image denoising method based on deep convolution neural network is proposed.In this method,the coefficients of transform domain are obtained by 3D shear wave transform,and the noise image is divided into two filtering stages: multi-scale decomposition and directional segmentation.Then,a 4D block matching method is proposed through the hard threshold and wiener filtering stages,including three steps of grouping,collaborative filtering and aggregation.The voxel cube stacked into 4D groups is used The 4D transformation makes use of the local correlation between voxels in each cube and the nonlocal correlation between voxels in different cubes.Through the inverse transformation of 3D shear wave,the estimated value of each group cube is obtained,and the adaptive aggregation is carried out at the original location,thus the potential clean image is obtained.Then we train the potential clean image with the generative adversarial networks of the deep convolution neural network,and get the final clean image.The method proposed in this paper can make full use of shear wave transform,and improve the 3D block matching algorithm,so that the 3D image can be denoised.Through the generative model and discriminative model in the deep confrontation learning network,the model is obtained through the confrontation training,and the fitting data distribution and data enhancement are realized.Compared with traditional denoising methods such as BM3 D,EPLL and tnrd,this papertakes the peak signal-to-noise ratio,structural similarity and edge retention index as the evaluation criteria to effectively improve the visual effect of the image.Moreover,compared with the traditional denoising methods,this method can effectively remove the image noise in the high noise environment,and proves that the method in this paper can effectively remove the image noise in the classic image and the magnetic field The effectiveness of denoising of vibration image.