Abstract: Image noise estimation is a very important research topic in digital image processing. This paper presents a fast and reliable noise estimation algorithm for additive white Gaussian noise. The proposed algorithm provides a way to measure the degree of image feature based on Statistical Hypothesis Tests（SHT）. The proposed algorithm distinguishes homogeneous blocks and non-homogeneous blocks by the degree of image feature. It sets the minimal variance of these homogeneous blocks as a reference variance. And then it finds more homogeneous blocks whose variances are similar to the reference variance and which not contain edge. The noise variance is estimated from these homogeneous blocks by a weighted averaging process according to the degree of image feature. Compared with the existing noise estimation methods, the proposed algorithm performs well in the estimation precision and suitable for the Gaussian noise-infected images.

Key words: white Gaussian noise, noisy image, noise estimation, statistical hypothesis tests

摘要： 在数字图像处理中，噪声方差估计是一个重要的研究课题。提出一种针对加性高斯噪声的噪声方差估计方法。利用一种基于统计假设测试的方法来度量图像结构特征度，基于图像结构特征度找出平滑子块和非平滑子块（含有边缘或纹理子块）；以平滑子块中的最小方差为参考方差，选择出方差与参考方差相差在一定范围内且不含边缘的所有子块；从选出的子块中求以图像结构特征度为权重的方差平均值作为噪声方差估计值。相比于现有的噪声估计方法，该方法具有非常高的估计精度，适合感染高斯噪声的各种图像。

关键词: 白高斯噪声, 噪声图像, 噪声估计, 统计假设测试