基于梯度特征的双核非局部均值去噪算法  

Dual kernel non-local means denoising algorithm based on gradient feature

作　　者：张玉征 杨词慧[1,2] 林泉 Zhang Yuzheng;Yang Cihui;Lin Quan(School of Information Engineering, Nanchang Hangkong University, Nanchang 330063, China;Key Laboratory of Jiangxi Province for Image Processing & Pattern Recognition, Nanchang 330063, China)

机构地区：[1]南昌航空大学信息工程学院,南昌330063 [2]江西省图像处理与模式识别重点实验室,南昌330063

出　　处：《计算机应用研究》2019年第5期1573-1576,共4页Application Research of Computers

基　　金：国家青年自然科学基金资助项目(61402218);江西省图像处理与模式识别重点实验室开放基金资助项目(TX201304002);南昌航空大学研究生创新基金资助项目(YC2016043)

摘　　要：针对传统非局部均值(NLM)滤波算法中邻域间相似性计算易受噪声干扰的问题,提出了一种基于梯度特征的双核非局部均值滤波算法。通过图像块之间的欧氏距离及梯度特征度量邻域间相似性,采用双核函数代替传统指数核函数计算相似性权值,并通过衡量搜索区域中的邻域块与当前像素邻域的相似程度,对像素点的权值进行重分配,在此基础上重估像素点去噪值并得到滤波图像。实验结果表明,提出的滤波算法与传统的NLM滤波算法及分别含有高斯核和正弦核的改进NLM滤波算法相比,可以更准确地反映邻域间的相似度,保存图像的细节及边缘信息,从而有效提升图像的去噪效果。In the traditional non-local mean (NLM) filtering algorithm,the presence of noise in the image interferes the accuracy of similarity calculation between neighborhood blocks.To address this problem,this paper proposed a dual kernel non-local means denoising algorithm based on gradient feature. This algorithm calculated the similarity of neighborhood block by Euclidean distance and gradient feature between pixels,and replaced the original exponential function with dual kernel function to calculate similar weights.In addition,it measured the similarity between the neighborhood blocks of the pixels in the search area and the current neighborhood to reassign the pixels’ weight.On this basis,it reevaluated the denoising values and got the denoised image.Experimental results show that this algorithm can accurately reflect the similarity between neighborhood block of pixels and preserve the details and edge information of images effectively while compared with the traditional NLM filtering algorithm and two other improved NLM algorithms,and improves the effect of image denoising significantly.

关 键 词：非局部均值 高斯函数 正弦函数 梯度特征 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]