Abstract In this paper, we propose a new texture-based conditional random field (CRF) for Synthetic Aperture Radar (SAR) image segmentation. In our proposed algorithm to overcome the limitations of the intensity-based features, feature extraction is performed in the contourlet transform domain. We use the nonsubsampled contourlet transform (NSCT) as an overcomplete transform which compensates the shortcomings of the traditional contourlet. Applying the generalized Gaussian distribution (GGD) for the statistical description of NSCT coefficients, we simultaneously extract proper statistics from SAR image in the conditional random field model and overcome the speckle effects in the intensity-based features. In this way, not only there is no need to consider an additional term in unary function to model the statistics of SAR image but also, we no longer need to calculate the several criteria based on the histogram of speckled gray levels. Experimental results show the superiority of NSCT compared to the other transform-based features such as wavelet and also demonstrate the improvement of the accuracy in contrast to the schemes which are based on the intensity in the CRF model.

中文翻译：

---

用于 SAR 图像分割的基于非下采样 Contourlet 变换的条件随机场

摘要 在本文中，我们为合成孔径雷达 (SAR) 图像分割提出了一种新的基于纹理的条件随机场 (CRF)。在我们提出的克服基于强度特征的局限性的算法中，特征提取是在轮廓波变换域中进行的。我们使用非下采样Contourlet 变换（NSCT）作为一种过完备变换，弥补了传统Contourlet 的缺点。将广义高斯分布 (GGD) 应用于 NSCT 系数的统计描述，我们同时从条件随机场模型中的 SAR 图像中提取适当的统计数据，并克服基于强度的特征中的散斑效应。这样，不仅不需要考虑一元函数中的附加项来对SAR图像的统计建模，而且，我们不再需要根据斑点灰度直方图计算几个标准。实验结果表明，NSCT 与其他基于变换的特征（如小波）相比具有优越性，并且与基于 CRF 模型中强度的方案相比，也证明了精度的提高。