基于光谱和形状的遥感图像分割质量评估方法

针对多尺度遥感图像的分割质量评估问题,提出了一种光谱和形状相结合的分割质量评估方法。首先,采用超像元方法对图像进行初始分割,将图像过分割为若干区域;其次,根据合并准则迭代合并相邻区域来生成各尺度图像,其中,使用尺度集结构来索引各尺度的区域,使用邻接图来记录各尺度下区域间关系;然后给出各尺度图像形状紧凑性和平滑性的计算公式,并结合各尺度图像光谱特征计算出各尺度图像的同质性和异质性;最后根据贝叶斯风险最小准则选择最优分割尺度。实验结果表明,该方法可以适应不同图像内对象特质,使得最优分割尺度的选择更合理,图像分割效果更佳。

Remote Sensing Image Segmentation Quality Assessment Based on Spectrum and Shape

With the continuous improvement of the spatial resolution of high-altitude platform remote sensing images, multi-scale information extraction methods have been used widely. The choice of optimal segmentation scale is a key technique in multi-scale remote sensing image segmentation. Aiming at the problem of segmentation quality assessment of multi-scale remote sensing image segmentation, a spectral and shape-based segmentation quality assessment method is proposed. Firstly, the image is initially segmented using the superpixel method, and the image is over-segmentated into several regions. Secondly, the multi-scale images are generated by iteratively merging the neighboring regions according to the merging criterion, and the scale-sets structure is used to index the regions of each scale. Adjacent graphs are used to record the relationship between the regions of each scale, then the formulas for the shape compactness and the shape smoothness of each scale are given by this paper, and the homogeneity and heterogeneity of each scale are calculated by combining the spectral features and the shape features of each scale; Finally, the optimal segmentation scale is automatically selected according to the Bayesian minimum risk criterion. The experimental results show that this method can adapt to the characteristics of regions in different images, and make the choice of optimal segmentation scale more reasonable and the image segmentation effect much better. The proposed algorithm selects the optimal segmentation scale based on standard of the global optimize, It is one of the development of multi-scale information extraction technology that how to make the object reach the best in every scale.