SAR image unsupervised segmentation using edge geometric complexity penalty
Wenjie Xu, Zejun Zhang, Jiabin Hu, Chenxi Min, Jianxiao Xie · 2025
In the region merging based image segmentation, the quality of final segmentation results highly depends on the order of merging adjacent region pairs. This letter proposes using geometrical edge penalty (GEP) to improve hierarchical region merging for segmenting SAR images with complex scenes, in which the boundary of each region is approximatively represented by straight segments head to tail and the GEP is constructed based on the numbers of segments and the pixels on the common boundary between two adjacent regions. An improved merging criterion is obtained by fusing the GEP with nonparametric Kuiper’s distance. Due to the existence of the GEP, adjacent regions with tortuous common boundaries and in the interior of a complicated region have a priority in merging order. Experimental results testify the effectiveness of the GEP and show that the improved method outperforms several published methods for SAR images with complex scenes.