DeepLabV3-SAM: A Novel Image Segmentation Method for Rail Transportation

Image segmentation is vital for improving visual perception in rail transportation. However, challenges like limited public data and high training costs exist. This study introduces the DeepLabV3-SAM model, combining DeepLabV3 and Segment Anything Model (SAM) for innovative image segmentation. By leveraging SAM's zero-shot transfer capability and merging it with traditional algorithms' semantic segmentation proficiency, our method achieves precise and fully automated segmentation with low training costs. This approach not only provides a groundbreaking solution to image segmentation challenges in railway transportation but also extends SAM's application by adding semantic segmentation capabilities. Performance evaluation on the RailSem19 dataset, a publicly available resource for semantic rail scene understanding, shows our method surpasses traditional models like DeepLabV3 and UNet, particularly excelling in segments with large-scale objects and well-defined boundaries. In the majority of cases, our DeepLab-SAM model outperforms traditional models, achieving a 10% lead in the mIoU metric across more than half of the segmentation tasks. Our method excels in achieving high-precision image segmentation with limited training samples and lower costs, offering a novel solution to segmentation challenges in rail transportation and related fields.