基于频率和张量投票的图像去噪及仿真研究  

Simulation Research of Image Denoising Based on Frequency and Tensor Voting

作　　者：柳婵娟[1,2] 钱旭[1] 厉彩霞[2] 

机构地区：[1]中国矿业大学机电与信息工程学院,北京100083 [2]鲁东大学信息与电气工程学院,烟台264025

出　　处：《系统仿真学报》2013年第2期333-339,345,共8页Journal of System Simulation

基　　金：国家自然科学基金(61170161);山东省自然科学基金(ZR2012FQ029)

摘　　要：结合人的视觉原理,基于图像局部梯度定义了一种图像频率,以图像频率代替全变分模型中的梯度。同时,把张量投票原理引入全变分模型,构造了一个图像结构显著性函数,代替变分模型中的拉格朗日乘子,根据图像不同区域的结构特征,去调节变分模型中正则项和保真项的作用,建立了一种基于频率的张量投票与全变分能量最小化结合的纹理图像去噪新方法。仿真实验数据表明,该模型比其他已有的全变分模型具有明显的抗干扰能力,能够更准确、精细地刻画图像边缘、特征结构和平滑区域,克服了其他全变分模型所产生的阶梯效应和过平滑现象,特别是对于纹理特征丰富和低信噪比的图像,在去除噪声的同时能较好地保持边缘及其他重要特征。Combining with human vision principle, firstly, an image frequency based on image local gradient was defined in the purpose of replacing the gradient in the traditional total variation （TV） model, and then tensor voting principle was introduced into the TV model and an image structure saliency function was given to replace the Lagrange multiplier;~, which could adjust the regularizing term and fidelity term according to the different areas of image structure features, therefore, a novel texture image denoising model integrating tensor voting and total variation minimization was constructed. Its simulation experiment results show that, compared with other existing TV approaches, the new model has an obvious anti- jamming capability and can accurately and subtly describe the sharp edges, feature structures and smooth areas, and can overcome staircase effect and over-smoothing generated by other TV models. Especially for those images with rich texture features and low signal to noise ratio （SNR）, it can remove the noise while preserving significant image details and important characteristics and improve the image denoising effect.

关 键 词：偏微分方程 全变分 张量投票 图像频率 纹理图像去噪 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]