A Research On Multi-spectrum Image Denoising Based On Tensor Regularized

Compared with traditional remote sensing images,multi-spectral remote sensing images can acquire more spectral information,thus creating more favorable conditions for the identification of boundaries and features,and therefore have higher research significance and application value.With the rapid improvement of spatial resolution in multi-spectral images in GIS technology,the processing and application of multi-spectral images have gradually increased.However,noise is inevitably introduced during the acquisition of multi-spectral images,affecting the sharpness of the image,which may cause difficulties for subsequent image processing and analysis.The current image de-noising methods mostly de-noise individual channels of multi-spectral images,but cannot effectively use the relevant information on multi-spectral images,often the denoising effect is not good,or new noise points are introduced.In this paper,based on the framework of tensor regularization,the noise reduction and image reconstruction of remote sensing multispectral images are studied.The tensor model is used to model the correlation between different channels.The main content and innovation of this article can be summarized as follows:1.Experimental analyses on traditional image denoising methods are performed to denoise the original multispectral image with Gaussian,salt,or speckle noise added.By analyzing the results of different noise sources and different denoising methods,the differences in denoising results of different denoising models for different noises are compared.The results show that the traditional denoising methods cannot satisfy the existing denoising expectations of multispectral images,and some of the details such as the edge can be lost.Meanwhile some image denoising methods such as Gaussian filtering and bilateral Butterworth filtering introduce new artifacts and cause image distortion problems.2.The tensor regularization framework was introduced into the image denoising model to improve the existing image denoising model.Through the data fitting and the regularization framework,the noise mechanism of the image and the prior information of the image are modeled respectively,so as to propose a stable and systematic image denoising modeling framework from a new perspective.The low-rank hypothesis is taken as a priori,and the low-rank regularization term is introduced.The Poisson noise(? = 6)and Gaussian noise(? = 0.2)were added to the original image.Simulationexperiments were performed on the model to recover the noise image and objectively evaluate the recovered image to verify the validity of the model.In addition,this paper introduces the current learning model into the regularization framework and combines it with the prioritization of low rankings to further improve the image denoising effect.3.Based on the non-local similarity of the multi-spectral image in the space and the global correlation in the whole spectrum,the dictionary learning is transformed into a new regular term and introduced into the image denoising model,and the building block similarity and sparsely constrained Dictionary Learning Based Tensor Regularization Learning Model is constructed.In this way,the same full frequency band is constrained by the spatially shared spectral lexicon.In addition,by using the spectral correlation of multi-spectral images and the constraints of excessive redundancy of the hypothetical dictionary,the restricted non-local multispectral image dictionary learning model can be decomposed into a series of easily solved low-rank constraints to approximate the solution of the problem.This paper uses traditional evaluation criteria such as peak signal-to-noise ratio(PSNR)calculation and runtime to evaluate the recovered image.Experimental results show that the proposed method can effectively improve these metrics,and subjectively,the image denoising effect is good.