Context-Aware Multi-Scale Feature Framework for Road Damage Detection under Complex Background Conditions With Heterogeneous Damage Patterns

Resumen

Automated road-damage detection is essential for timely pavement maintenance and intelligent infrastructure management; however, heterogeneous damage patterns and complex backgrounds containing shadows, vehicles, vegetation, lane markings, stains, and debris continue to challenge existing deep-learning detectors. Conventional approaches often struggle to preserve fine-grained features of small cracks, distinguish damage from visually similar backgrounds, and adapt feature representations to varying damage morphology. To address these limitations, this study proposes a Context-Aware Multi-Scale Feature Framework based on RT-DETRv2 with a novel Context-Adaptive Heterogeneous Damage Selection Algorithm (CA-HDSA). The proposed method first extracts multi-scale contextual representations using RT-DETRv2, after which CA-HDSA dynamically evaluates damage response, contextual relevance, background interference, and damage heterogeneity to prioritize informative features. The selected representations are subsequently integrated through adaptive multi-scale fusion and processed by RT-DETRv2 for object-level classification and localization of cracks, potholes, and surface erosion. The framework is implemented using Python and PyTorch and evaluated on a curated RGB road-damage dataset comprising 1,075 annotated images, with RDD2022 employed for external validation. The illustrative experimental results indicate 96.18% precision, 95.64% recall, and 95.91% F1-score, with mAP@0.50 of 96.27% and mAP@0.50:0.95 of 72.86%. Compared with the RT-DETRv2 baseline, the proposed framework improves F1-score from 89.32% to 95.91%, representing a 7.38% relative improvement, while substantially reducing background-induced false positives. These findings indicate that CA-HDSA provides an effective mechanism for context-sensitive feature selection and robust multi-scale road-damage detection, offering a promising foundation for scalable intelligent pavement inspection systems.