两类噪声背景下的非局部图像去噪研究

项目名称： 两类噪声背景下的非局部图像去噪研究

项目编号： No.61471004

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 无线电电子学、电信技术

项目作者： 许光宇

作者单位： 安徽理工大学

项目金额： 80万元

中文摘要： 噪声的存在严重影响图像的视觉质量和后继处理。本项目研究高斯噪声和脉冲噪声背景下的非局部图像去噪方法。对于高斯噪声，利用梯度域奇异值分解研究基于预选择的非局部均值图像去噪方法，从去噪效果和运行速度两个方面对原算法进行改进；研究有效的像素相似性度量方法，包括两个部分：1）旋转匹配图像片相似性度量；2）方向相关相似距离计算，该方法能够获得较多的相似像素和鲁棒的相似权系数。对于脉冲噪声，采用极值压缩顺序阶绝对差统计方法检测图像脉冲像素，联合噪声检测结果与非局部均值滤波框架设计一种通用的图像脉冲噪声去除方法（椒盐噪声、随机脉冲噪声和混合噪声）。构造的脉冲权可以避免恢复过程中噪声像素的影响，非局部信息能够在噪声像素和其邻域像素之间提供较高的相关性，有助于噪声抑制和边缘保持。研究非局部约束的最优化方法，给出三种图像去噪模型。

中文关键词： 图像去噪；非局部信息；相似性度量；通用脉冲噪声滤波器；最优化方法

英文摘要： Noise seriously affects image visual quality and subsequent processing operations. The research content of this project is the nonlocal based image denoising methods for Gaussian noise and impulse noise. For Gaussian noise,the preselection based nonlocal means (NLmeans) denoising method is studied by employing singular value decomposition in the image gradient domain, and the traditional NLmeans method is improved in denoising effectiveness and running speed. We study an effective similarity measure method between pixels, which consists of two parts: 1) Rotation matching based image patch similarity measure; 2) Direction correlation based similarity distance computation. Such similarity measure can obtain more similar pixels and robust similar weights. For impulse noise, a local image statistic, called the extrumum compression rank-order absolute difference (ECROAD), is employed to detect impulse noise in an image. A universal impulse noise removal method, used to remove salt-and-pepper noise, random valued impulse noise and mixed impulse noise, is designed by combining the ECROAD statistic results with NLmeans filtering framework. The designed impulse weight is able to avoid the effect of noisy pixels in restoring candidates. The nonlocal information can provide higher correlation between the corrupted pixel and neighborhood pixel. Higher correlation gives rise to better noise suppression and edge preservation. We study optimization method based on nonlocal constraint and present three different image denoising models.

英文关键词： Image denoising;Nonlocal information;Similarity measure;Universal impulse noise filter;Optimization method