基于最小特征值非线性修正的快速噪声水平估计算法  

Fast Noise Level Estimation Algorithm Based on Nonlinear Rectification of Smallest Eigenvalue

作　　者：徐少平[1] 曾小霞 姜尹楠 林官喜 唐祎玲[1] XU Shao-ping;ZENG Xiao-xia;JIANG Yin-nan;LIN Guan-xi;TANG Yi-ling(School of Information Engineering,Nanchang University,Nanchang 330031,China)

机构地区：[1]南昌大学信息工程学院,南昌330031

出　　处：《计算机科学》2018年第7期219-225,共7页Computer Science

基　　金：国家自然科学基金(61662044;61163023;51765042;81501560);江西省自然科学基金(20171BAB202017)资助

摘　　要：鉴于从噪声图像上提取的原生图块协方差矩阵的最小特征值与噪声水平值之间具有显著的相关性,提出一种基于多项式回归技术训练非线性映射模型,直接将原生图块最小特征值修正为最终的噪声水平预测值的快速噪声水平估计算法。首先,选择具有代表性且无失真的自然图像作为训练图像集合;然后,对这些图像施以不同程度的高斯噪声构成样本训练图像库。在此基础上,提取各个噪声样本图像的原生图块,并使用PCA变化得到原生图块协方差矩阵的最小特征值;最后,利用多项式回归技术构建最小特征值与噪声水平值之间的非线性修正模型。实验表明,与现有算法相比,改进算法对高、中、低各级别的噪声都能鲁棒地进行预测,尤其在低水平噪声方面表现出色,在预测准确度和执行效率两方面具有显著的综合优势。Considering the fact that the smallest eigenvalue of covariance matrix of the raw patches extracted from noise images is significantly correlated with noise level,this paper proposed a fast algorithm that directly uses a pretrained nonlinear mapping model based on the polynomial regression to map（rectify）the smallest eigenvalue to the final estimate.Firstly,some representative natural images without distortion are selected as training set.Then,the training sample library is formed,and the training set images are corrupted with the different noise levels.Based on this,raw patches are extracted for each noisy image,and the smallest eigenvalue of covariance matrix of the raw patches is gotten by PCA transformation.Finally,a nonlinear mapping model between the smallest eigenvalue and the noise level are obtained based on polynomial regression technique.Extensive experiments show that the proposed algorithm works well for a wide range of noise levels and has outstanding performance at low levels in particular compared with the existing algorithms,showing agood compromise between speed and accuracy in general.

关 键 词：图像降噪 噪声水平估计 主成分分析 最小特征值 修正函数 低水平噪声 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]