为了减少仅用分水岭变换而导致的过分割问题，本文提出利用小波变换的多尺度处理方式用于融合后多光谱QuickBird图像的分割。整个分割过程包括多尺度图像表示、图像分割、区域合并和结果映射等过程。首先，依据原始图像的大小确定分解尺度并用小波变换产生各波段的低尺度图像。采用相位一致模型提取各近似系数的梯度，并逐尺度地融合各梯度图。分析不同尺度下的不同地物的局部梯度方差，以选择最佳的小波分解尺度。然后，通过移动阈值与扩展最小变换，利用多层次标记提取方法标记均质区域。进而，在梯度重建的基础上利用标记分水岭变换得到分割图像。其次，采取空间相邻关系、面积、光谱与纹理等多约束策略，以搜索最小合并代价的方式合并最初分割区域中的邻接区域对。最后，修改细节子图并进行小波逆变换将最初分割结果投影到更高尺度图像，同时处理边界上的像元以保持区域边界直至原始图像。实验结果表明本文方法不仅能够用于高分辨率多光谱遥感图像的分割，而且缓解了过分割问题且取得了较准确的分割效果。

In order to reduce over segmentation caused by only using watershed algorithm, an effi cient multi-scale approach using wavelet transform is presented for the segmentation of the pan-sharpened multi-spectral QuickBird image. The approach toward complete segmentation includes four steps, namely, multi-scale images representation, image segmentation, region merging and result projection. First, the wavelet decomposition is implemented independently for each spectral band to form a number of new images at lower resolutions according to the size of original image. Gradient images are obtained by applying phase congruency model to approximation coeffi cients, and gradient magnitudes of all bands are fused for each scale. The optimal scale of wavelet decomposition is selected through analysis local gradient variance varying correspond to different scales and varieties of geo-objects. Second, a multi-level marker extraction algorithm is subsequently used to locate regions that are homogeneous, by moving threshold and extended minima transform. A marker driven watershed transform is then used to get segmented image based on gradient reconstruction. Third, a multi-constraint region merging strategy considering spatial adjacency relation, areas, spectral and textural features is proposed to merge the adjacency region pairs by searching the minimum merging cost among the initial segments. Finally, the detail coeffi cients are updated and the inverse wavelet transform is computed to project the initial segmentation to higher scale images, and pixels at boundaries are processed to keep region contours as original scale is reached. The experimental results demonstrate that the developed method can be applied to the segmentation of high-resolution multispectral remote sensing image as well as alleviate over segmentation and get the high accuracy segmentation.