Image Denoising Based On Multi-scale Geometric Decomposition And Neural Network

With the development of computer technology and science,a variety of imaging equipment and imaging methods have made great progress,and are widely used in various fields,making image processing technology increasingly important.The distortion of images corrupted by noise is common during its acquisition,compression,storage and transmission,images noise often leads to the loss of important details.In addition,due to the fact that the ground truth,noise type and noise density level of the real noisy image are unknown,the denoising of real images is a great challenge.In this paper,a new method of multi-scale geometric decomposition and neural network is used to denoise images.Compared with traditional methods,the denoising image processed by the new method we proposed retains more detailed texture information.The research content of this paper mainly includes the following three aspects:(1)Modern camera equipment produces a large number of images with noise.Removing noise can give full use to the utilization value of images,but at the same time,image edge and texture information will be destroyed.To solve this problem,we proposed an image denoising algorithm based on the multi-scale geometric transform combine with Fuzzy Support Vector Machine(FSVM).First,the multi-scale geometric transformation is used to decompose the image,and a binary coordinate map is constructed for the decomposed subband coefficients,and spatial rules are used to judge whether the coefficients contain the image feature information or not.After the training of FSVM,the coefficients of each high-frequency subband are divided into detail class and noise class,and a justifiable granularity method is proposed to solve the parameter selection problem.According to the adaptive Bayesian threshold,the corresponding threshold value of each subband are calculated,and the coefficient shrinkage of each subband is carried out.The proposed algorithm is suitable for denoising images and can improve the quality of denoising images.Compared with the results of similar denoising algorithms,the effectiveness of the algorithm is verified.(2)Most deep convolutional denoising networks lack model flexibility and the trained denoising models can only deal images with specific noise density levels.To solve this problem,a denoising model of convolutional neural networks based on Discrete Shearlet Transform is proposed.Using Discrete Shearlet Transform to decompose images,the number of subband image is obtained by decomposition as the training sample input CNN module,this operation expands the scale of the sample,relieves the problem of small sample model fitting trained and improve the generalization ability,to reduce the time needed for model training.At the same time,the Discrete Shearlet Transform can pre-classify the image information and noise,which helps the model to better extract image features and improve the denoising performance of the model.In the end,the inverse Discrete Shearlet Transform algorithm is used to reconstruct the image.Since the Discrete Shearlet Transform has the property of lossless information,it will not cause the loss of texture of the image,which is helpful to preserve the details of the image.The proposed method uses a variety of strategies in the training stage and introduces a random noise graph so that the model has sufficient flexibility to denoise the images with different noise density levels.It is suitable for the real image denoising and improves the flexibility of the model on the premise of ensuring the denoising effect.(3)To further enhance the denoising ability of the deep convolution neural network model,better keep the edge of the denoising image texture information,solve the problem of dilated convolution.A Nonsubsampled Countourlet Transform(NSCT)based deep convolution denoising neural network is proposed.The network model consists of a contraction subnetwork and an expansion of subnetwork,thanks to the excellent time and frequency domain characteristics of NSCT,NSCT and inverse NSCT are used to instead of sum pooling layer and up-convolution operation,which improving the denoising effect of the model,no the information loss and dilated convolution are avoided.Compared with similar denoising algorithms,the proposed model has excellent denoising effect,moderate computational complexity,and can deal with range noise.It is suitable for denoising actual images and has a broad application prospect.