Research And Implementation Of Color Image Denoising Based On Gaussian Weight And Manifolds For High Fidelity

With the popularization of digital camera and other color image acquisition device,the color image contains more information than of gray-scale images. People tend tochoose color image as the main means of access to information. Nevertheless, colorimages are often corrupted by noise during acquisition, transmission, processing,storage and other processing. And affect the subsequent image processing. The mainpurpose of the color image denoising is to as far as possible keep texture informationand color information of original image, the image structure cannot be destroyed at thesame time of eliminate the noise.In order to realize real-time color image denoising. This thesis use the method ofimages denoising algorithm for gray image and the high-dimensional filter. Improvedthe algorithm, can better keep the details of the color image texture. At the same time, itwork on color images denoising at arbitrary scales in real time. The main researchcontents in this article are as follows:①To Gaussian random nosie removal in color images, inspired by non-localmeans filter for gray-scale images denoising, a improved non-local means algorithm isproposed. Firstly, introduces the traditional non-local means filter for gray image noisereduction processing. Then, according to the structure characteristics of color images,improved non-local means filter which can implementation of color image noisereduction processing is proposed. Secondly, based on other scholars research on thekernel function of non-local means filter. Improved the kernel function of non-localmeans filter for color images noise reduction. Lastly, we evaluate and compare thedifferent kernel function by using two measures: PSNR and SSIM.②To improve the effect of the color image noise reduction processing, imagedenoising by principal component analysis with local pixel grouping is proposed. Firstly,modeling of spatially adaptive PCA denoising, extract the training sample. Secondly,calculate the PCA covariance matrix, using linear minimum mean square erroralgorithm for noise reduction processing. Thirdly, implement the LPG-PCA denoisiongprocedure for the second round to enhance the denoising results. Lastly, experimentalresults and analysis has been given.③In order to solve the real-time problem of color images denoising, combinedwith the non-local means filter and high dimensional filter, we proposed the algorithm of color image denoising based on Gaussian weight and manifolds. Our method hasfour step: First, using the non-local means algorithm to get high-dimensional data, thenon-local means use the improved Gaussian kernel; Secondly, splatting performs aGaussian distance-weighted projection of the colors of all pixels onto each adaptivemanifold; Thirdly, blurring performs Gaussian filtering over each adaptive manifold,mixing the splatted values from all sampling points; Finally, slicing computes the finalfilter response for each pixel by interpolating blurred values gathered from all adaptivemanifolds. Experimental results show that the algorithm has a superior denoisingperformance than the original one, the details and images structure can be preservedwell.