EATP-net: An edge-aware and topology-preserving network for instance segmentation of underground drainage pipe defects
DOI number:
10.1016/j.jwpe.2026.110699
Journal:
Journal of Water Process Engineering
Key Words:
Instance segmentation; Underground drainage pipe; Edge-aware; Topology preserving; Multi-scale feature fusion
Abstract:
Accurate instance segmentation of underground pipeline defects is crucial for structural assessment. However, conventional models deployed in subterranean environments suffer from high-frequency gradient degradation, perspective-induced scale distortions, and boundary ambiguity under low signal-to-noise ratios. To resolve these bottlenecks, this paper proposes the Edge-Aware and Topology-Preserving Network (EATP-Net). First, a Global Edge Information Transfer (GEIT) backbone isolates and preserves fragile topological priors against spatial quantization errors. Second, a Gather-and-Distribute with Attentional Scale Sequence Fusion (Gold-ASF) neck replaces recursive feature pyramids to establish robust affine-invariant perception across varying defect scales. Finally, a Lightweight Shared Convolution Detection with Localization Quality Estimation (LSCD-LQE) head introduces quality-aware prediction to filter deceptive background artifacts. Crucially, LSCD-LQE acts as an efficiency compensator, offsetting the computational overhead introduced by the structural enhancements. Extensive evaluations on a multi-source dataset demonstrate EATP-Net's superiority. Operating at a moderate 12.2 GFLOPs, it achieves a Mask mAP50 of 0.836, successfully balancing topological fidelity with edge-deployment feasibility. Notably, it yields up to 16% absolute accuracy improvements on challenging anomalies like fine ruptures and root intrusions, confirming its robustness for real-world automated robotic inspections.
Paper Publications
