A Few-Shot Road Extraction Method Using Customized Segment Anything Model

With the rapid development of drone technology, unmanned aerial vehicle (UAV) technology has become a promising tool for monitoring the state of transportation infrastructure. Deep learning-based semantic segmentation algorithms for extracting road surfaces have shown promising performance. However, the utilization of supervised learning techniques requires the labor-intensive collection of datasets and extensive pixel-level annotation. Therefore, we propose a few-shot segmentation program based on the Segment Anything Model (SAM) for road extraction in high-resolution drone images. This method employs Self-Distillation with No Label Version 2 (DINOv2) and the clustering analysis to select representative images. On the one hand, positive prompts are generated by multiple SAM models based on representative images. On the other hand, negative prompts are produced using You Only Look Once-World (YOLO-World) with text prompts. Experimental results indicate that the proposed method reaches 90.89 % intersection over union and 94.79 % accuracy in road extraction on UAV images. Compared with UNet, our method has better capability in segmenting the general road contours.