Hierarchical Multiscale Fusion Enhanced Deep Bi-resolution Network for Efficient Semantic Segmentation of Railroad Scenes

In the context of intelligent transportation and autonomous driving, real-time semantic segmentation of complex railroad scenes is of critical importance for safety decision-making. Considering that there is a trade-off between accuracy and inference speed, and that there are still challenges in recognizing small objects and complex textures, a method combining the Hierarchical Multi-scale Fusion Module with the Deep Dual-Resolution Network is proposed. DDRNet uses parallel high-resolution and low-resolution branches, with the high-resolution branch preserving spatial details and the low-resolution branch extracting deep semantic features, and adopts bidirectional feature fusion to improve cross-resolution interactions, while the HMSFM module optimizes contextual modeling by hierarchically aggregating multiscale features on the low-resolution branch. The semantic representation of small targets such as railroad boundaries and signaling equipment is enhanced. Experimental results on the railroad scene dataset Railsem19 show that the method achieves an excellent balance of 63.75% mIoU and 106.15 FPS, which outperforms current ERFNet and BiSeNetV1 real-time models, and verifies its robustness to classify complex railroad elements with a Macc of 74.30%, making the architecture a perfect solution for the classification of complex railroad elements in challenging railway environments. This architecture provides an effective solution for real-time semantic segmentation in challenging railroad environments and promotes the development process of real-time semantic segmentation for complex railroad scenarios.