Road-oriented attention and class-balanced learning for post-disaster road damage segmentation

Engineering notes

On the held-out test set, the proposed framework achieves an mIoU of 0.7248 and a damage-class IoU of 0.6029, outperforming representative segmentation baselines including U-Net, FCN, ResUNet, standard DeepLabV3+, and MLA DeepLabV3+.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Accurate pixel-level segmentation of post-disaster road damage is important for decision-support applications that require reliable and spatially explicit road-scene understanding. This study proposes a deep-learning framework for localizing visible road-damage regions in post-disaster imagery. The framework integrates a Road-Oriented Multi-Level Attention (RO-MLA) module into a DeepLabV3+ backbone, where channel-, spatial-, and pixel-level attention are guided by an average road-mask prior to concentrate feature learning on road-relevant regions. To address severe class imbalance and weak feature contrast in disaster-related road imagery, we further introduce a Road-Oriented Class-Weighted Balanced Dice (RO-Dice CWB) loss to improve the segmentation of sparse, low-contrast, and underrepresented damage regions. To evaluate the framework, we construct a merged dataset from the Post-Disaster Road Damage Dataset-Japan (PDRDD-J) and the Social Media Disaster Road Damage (SoDR) dataset, yielding 2684 pixel-level annotated images across real-world road-damage scenarios. Validation-based ablation studies support the selection of the DeepLabV3+ base architecture, ResNet152 backbone, and RO-Dice CWB loss. On the held-out test set, the proposed framework achieves an mIoU of 0.7248 and a damage-class IoU of 0.6029, outperforming representative segmentation baselines including U-Net, FCN, ResUNet, standard DeepLabV3+, and MLA DeepLabV3+. Qualitative, resolution-sensitivity, and error-map analyses further show that the proposed attention and loss design produces more coherent damage masks while maintaining stable road-region interpretation. These results demonstrate that the proposed framework supports post-disaster road damage segmentation and offers a task-specific approach for learning sparse, imbalanced, and visually ambiguous targets in complex road scenes.

Links and sources

• PDF from original source

Comments