基于最小核值相似区算法的高分辨率遥感图像分割方法

Segmentation of the High Spatial Resolution Remotely Sensed Imagery Based on SUSAN

采用最小核值相似区算法(Smallest Univalue Segment Assimilating Nucleus,SUSAN)计算QuickBird图像的梯度,并采用标记控制的分水岭变换(Watershed Transform,WT)算法分割图像,取得了较好的结果。SUSAN方法能有效地检测图像梯度,对噪声不敏感;梯度值范围明确,不因图像而改变,为后续处理相关参数的选择提供了便利;亮度阈值容易确定,模板半径可选,具有很大的灵活性;适合于采用WT的遥感图像的分割。采用基于SUSAN梯度和NDVI的标记图像,利用形态学灰度图像重建方法修改梯度图像,能够有效地抑制梯度图像中大量的局部灰度极小值,提高WT图像分割的精度。

The SUSAN（Smallest Univalue Segment Assimilating Nucleus） method is used to detect gradient features from QuickBird imagery,and then the imagery is segmented using marker-controlled WT（Watershed Transform）,and the segmentation result is satisfactory.The SUSAN method detects gradients well.It is not sensitive to noise and the values of the gradients are in a definite range and do not change with images,which offers convenience in selecting parameters in the later processes.The method is flexible because it is easy to choose the illumination threshold and the size of SUSAN matrix is not fixed.Based on the marker derived from both SUSAN gradients and NDVI,the gradients are modified using morphological grayscale reconstruction method,which efficiently constrains much local minima of the gradients and improves the segmentation precision.