分组无监督域适应的遥感影像道路提取

Grouped unsupervised domain adaptation for road extraction of remote sensing images

针对跨域使用道路语义分割模型时,道路提取精度往往会大幅度下降的问题,采用针对性策略研究了一种分组迁移的跨域遥感影像道路提取算法:基于语义一致性对源域影像和目标域影像进行聚类生成跨域影像组,增强用于风格迁移的组内影像的语义相似性;采用循环对抗生成网络对组内影像进行风格迁移,降低影像迁移难度并提高迁移的有效性;利用改进的跨域伪标签交叉监督算法进一步提高源域模型对目标域影像的道路提取精度。实验结果表明,该文算法能够更加准确的提取目标域影像中的道路信息,在DeepGlobe数据集上训练的模型在Massachusetts测试集上的交并比达到了56.96%,优于其他几种同类无监督域适应算法。

When applying road semantic segmentation models across different domains,the accuracy of road extraction often significantly decreases.This paper develops a targeted group transfer algorithm for cross-domain road extraction from remote sensing images.Initially,images from the source and target domains are clustered based on semantic consistency to form cross-domain image groups,enhancing semantic similarity within the groups for style transfer.Subsequently,a cycle-consistent adversarial network is employed to perform style transfer on images within these groups,reducing the difficulties associated with image transfer and enhancing its effectiveness.Finally,an enhanced cross-domain pseudo-label cross-supervision algorithm is utilized to further improve the road extraction accuracy of the source domain model on target domain images.Experimental results demonstrate that our algorithm more accurately extracts road information in target domain images,and the model trained on the DeepGlobe dataset achieves an Intersection over Union(IoU)score of 56.96%on the Massachusetts test set,which surpasses other comparable unsupervised domain adaptation algorithms.