Answers are grounded in this paper and cite their sources.
Objectives To address the challenges posed by complex backgrounds, regional structural relationships and class imbalance in road distress recognition and localisation, this study develops a framework integrating structural region modelling and multi-scale query mechanisms.Methods For classification, visual features are partitioned into a 4×4 grid, with a region graph constructed from spatial adjacency and feature similarity. A graph convolutional network (GCN), bidirectional long short-term memory (BiLSTM) and gated attention capture local relations, cross-region context and global semantics. For localisation, multi-scale features, feature reweighting and learnable object queries predict distress classes and bounding boxes. Models are evaluated on an independent RDD-China test set after an image-level stratified 70:15:15 split.Findings Classification achieved 84.85% accuracy, 76.47% Macro-F1 and 84.73% Weighted-F1. Removing BiLSTM and GCN reduced Macro-F1 by 4.12 and 0.17 percentage points, respectively. Localisation achieved 57.20% mAP@[0.5:0.95], with 12.5 ms inference latency and 80.0 FPS. Paired bootstrap analysis supported a 5.03-percentage-point mAP improvement over YOLOv8s. Inference was faster than RT-DETR-L (12.5 versus 20.3 ms).Novelty The framework combines regional structural reasoning with multi-scale query-based localisation to balance predictive performance and inference efficiency.
No takes yet. Share an insight, caveat, or question.
Xunqian et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: