目的 结合高斯核函数特有的性质，提出一种基于结构相似度的自适应多尺度SAR图像变化检测算法。方法 本文提出的算法包括差异图像获取、高斯多尺度分解、基于结构相似性的最优尺度选择、特征矢量构造以及模糊C均值分类。首先，通过对多时相SAR图像进行对数比运算获取差异图像，然后，利用基于图像的结构相似度估计高斯多尺度变换的最优尺度，继而在该最优尺度参数下逐像素构建变化检测特征矢量，最后通过模糊C均值聚类方法实现变化像素与未变化像素的分离，生成最终的变化检测结果图。结果 在两组真实的SAR图像数据上测试本文算法，正确检测率分别达到0.9952和0.9623，Kappa系数分别为0.8200和0.8540，相比传统算法有了较大的提高。结论 本文算法充分利用了尺度信息，对噪声的鲁棒性有所提高。实测SAR数据的实验结果表明，本文算法可以智能获取最优分解尺度，显著提高了SAR图像变化检测性能。
Multi-scale approach based on structure similarity for change detection in SAR images
Objective Synthetic Aperture Radar (SAR) is suitable for dynamic monitoring because it is unaffected by weather conditions. SAR image change detection is the key technology for the dynamic monitoring of targets. However, multi-scale SAR images contain more details than single-scale images. Thus, we integrate multi-scale analysis in the change detection algorithm to obtain accurate results. Gaussian multi-scale theory is easy to understand and has advantages. Despite its capacity to preserve image details, it is seldom used in change detection. This study proposes a new adaptive multi-scale change detection method on the basis of the structure similarity (SSIM) of SAR images. Method The proposed method includes five steps, namely, obtaining difference images (DIs), multi-scale decomposition based on Gaussian kernel, optimal scale estimation by SSIM, building feature vectors for each pixel, and fuzzy C-means (FCM) clustering to obtain the final change map. DIs are obtained using a log-ratio operator. A median filter is utilized to suppress the speckle noise. The optimal Gaussian scale is estimated by finding the maximum SSIM of the DIs through iterations. The optimal Gaussian kernel scale and its differential forms are convoluted with the DIs to generate change detection feature vectors at the pixel level. FCM is introduced to classify the changed and unchanged pixels using the feature vectors, and the change detection map is achieved. Result Experiments on two pairs of real SAR images of Bern and Ottawa areas show that the proposed method outperforms state-of-the-art algorithms. The correct detection rates of the two pairs of SAR images reach 0.9952 and 0.9623, and their kappa coefficients reach 0.8200 and 0.8540. Conclusion This study proposes an effective multi-scale change detection method based on the SSIM of SAR images. This method fully utilizes the scale information of SAR images and is robust to speckle noise. The proposed method is effective in finding the optimal scale of DI and in detecting changes.