| Title |
A Three-Multihead-Based Framework for Simultaneous Multi-Damage Segmentation in Reinforced Concrete Structures |
| DOI |
https://doi.org/10.12652/Ksce.2026.46.4.0363 |
| Keywords |
철근콘크리트 구조물, 복합 손상 분할, Three-Multihead, 균열, 철근노출, 박락 Reinforced concrete structures, Multi-damage segmentation, Three-Multihead, Crack, Rebar exposure, Spalling |
| Abstract |
In this study, a Three-Multihead semantic segmentation framework was proposed for the simultaneous segmentation of crack, rebar exposure, and spalling in reinforced concrete structures. The proposed model incorporates Knowledge Distillation (KD), Pairwise Prior, and Extra Loss to improve composite-damage segmentation performance by transferring knowledge from single-damage teacher models, reflecting spatial interactions among damage classes, and suppressing overlapping activations. Experimental results showed that the proposed model achieved a 13.5 % improvement in mean Dice compared with the baseline model and a 20.0 % improvement compared with the Multi-Damage Representation model. The ablation study further demonstrated that Pairwise Prior and Extra Loss effectively reduced inter-class interference and overlapping predictions. To evaluate performance under real composite-damage conditions, a dataset consisting of 44 composite-damage images was constructed. Without additional training, the proposed model achieved a mean Dice score of 0.35, which increased to 0.57 after fine-tuning using the composite-damage dataset. Class-specific threshold optimization provided a modest calibration effect, resulting in a final mean Dice score of 0.58 while improving the prediction balance across damage classes. These results indicate that composite-damage data-based learning is effective for improving model adaptation to real composite damage environments and suggest that the proposed Three-Multihead framework is a promising approach for enhancing composite-damage segmentation performance. |