遥感图像扫描聚类分割算法

针对传统聚类分割算法对噪声和异常值敏感的问题，利用RGB特征空间对图像主体的表达能力及图像空间的邻域相关性，基于扫描聚类原理提出一种对噪声和异常值具有较强鲁棒性的遥感图像分割算法。算法以椭球形建模不同目标在RGB特征空间分布，以三个主成分方向逼近椭球三轴方向，通过逐渐调整椭球中心、主成分方向以及三轴长度使椭球对目标主体在RGB特征空间分布达到最优拟合效果，进而实现RGB特征空间不同目标主体的识别。RGB特征空间能够有效区分不同类型目标，但无法判定远离聚类主体像素的隶属性。为了充分利用图像空间相邻像素隶属于同一目标的可能性较大这一性质，利用相邻像素标号填补主体分割结果中的空洞区域，得到对噪声具有较强鲁棒性的完整分割结果。对合成遥感图像和真实遥感图像的分割实验表明，提出算法不但能够有效识别目标主体，还能极大程度地提高算法对噪声和异常值的鲁棒性。 

The Remote Sensing Image Scan Clustering Segmentation Algorithm

Aiming at solving the problems that traditional cluster segmentation algorithms are sensitive to noise and outliners, a remote sensing image segmentation algorithm was introduced based on scanning clustering theory according to the main body expression capacity of image of RGB eigenspace and neighbor correlation of image space, which has a higher level of noise and outliners robustness. This algorithm models the RGB eigenspace distribution of different targets using ellipsoidal objects. In the model, three principle component directions approach three axis of the ellipsoid. By adjusting the ellipsoid center, principle component direction and the length of three axis, the ellipsoid will be a best fit for target bodies in the distribution of RGB eigenspace. In this case, the recognition of different target bodies in RGB eigenspace will be achieved. RGB eigenspace is capable to distinguish different type of targets effectively. However, it cannot determine subordination of the body pixels away from clusters. For fully use of the characteristics that neighbor pixels are more likely to be subordinate to the same target in the image space, neighbor pixel symbols are used to fill the holes in the result of body segmentation for getting a complete segmentation result which has a better robustness of noise. According to the experiments of synthesis remote sensing images and truth remote sensing images, not only can the proposed algorithm differentiate body targets effectively, but also improve the noise and outliners robustness greatly.