Road Graph Extraction of Satellite Images Based on Graph Similarity
The road network in satellite images is a typical curved structure,and its extraction is an important problem in com-puter vision.For such a structure,topology is the most important feature and consistent connectivity between predictions and ground truth should be ensured.In the road network,the disconnection of a single road segment will completely change the navigation re-sult.Many extraction methods focus on using deeper neural networks,but still habitually use pixel-wise loss functions,which may-be not suitable for this problem due to the absence of topology supervision.In this paper,a connectivity-oriented loss is discussed.The main idea is to measure the connectivity of the road network mediately according to the separation caused by the road segment between the background regions.Obviously,if the extracted road is disconnected,the resulting gap will cause the connection of the background areas which should lie on both sides of the road.The loss function is designed to suppress such wrong connectivity be-tween background regions and thus suppress the disconnection of the road network.In addition,the loss function penalizes false dis-connections of the background region,thus reducing the generation of false positive examples in the predictions.It demonstrates this loss function improves the road network connectivity and it's enough to be skeletonize it to generate maps comparable to current most used networks.The loss function can be easily combined with any semantic segmentation network.
satellite imagesroad extractionconnectivity-oriented loss function
国家科技期刊平台
NSTL
万方数据
