Research Of Image Denoising Based On Partial Differential Equation And Non-local Means

The image is very susceptible to noise and becomes blurred during transmission.To obtain high-quality images,image denoising technology has become a hot research field in image processing.How to effectively remove noise and ensure that the structural information such as the edge and texture of the image is not destroyed has always been the main problem to be solved in image denoising.The image denoising method based on Partial Difference Equation(PDE)has developed rapidly.The construction of a suitable partial differential equation model is the key to denoising based on PDE.This article mainly studies the following two types of modeling methods,one is based on total variation(Total Variation,TV)denoising model;the other is direct Establish corresponding partial differential equation models for the evolution process.The traditional full variational denoising method will produce a staircase effect,and there is a contradiction between edge preservation and noise smoothing.In order to solve this problem,this dissertation improves Zhang Hongying's adaptive full variational model.Since the edges of natural images are relatively smooth curves,and the edges of images interfered by noise are oscillating,we will use the horizontal set curvature as The control variables are introduced into the regular term of the adaptive full variational model,and the gradient value is originally used in the model to change the curvature drive function to determine the adaptive value.Experiments show that this method effectively suppresses the step effect,alleviates the contradiction between noise removal and edge retention,and retains more image edge texture information than the original model,effectively improving the quality of the denoised image.The image itself has the characteristics of structural information redundancy and autocorrelation.Non-local means(NLM)achieves a good denoising effect by using the Euclidean distance of similar image blocks between pixels as similarity measure However,the algorithm relies heavily on the accuracy of the similarity measure,which is susceptible to noise and reduces accuracy,and the algorithm has high complexity.The Fast Non-local Means(FNLM)algorithm can greatly shorten the calculation time,but the denoising effect has not been improved.In order to solve this problem,this dissertation proposes the following two improvement methods.The NLM algorithm is actually a linear expression of the similarity weight between image blocks,and the BP neural network has a very powerful non-linear mapping ability.This article uses the BP neural network feature to improve the NLM algorithm and improve the effect of image denoising;Due to the redundancy and self-correlation of the image,most natural images are sparse,especially the method noise of the NLM algorithm is highly sparse,so this dissertation applies the sparse representation denoising method to the noise of the NLM algorithm method In the two-stage denoising framework,a good visual effect is obtained.