TunnelMIND: Training-Free Defect Inspection for Tunnel Engineering
A novel framework named TunnelMIND has been introduced for inspecting tunnel defects, as outlined in a preprint on arXiv (2604.27928). This system overcomes the challenges posed by current foundation-model pipelines that generate imprecise open-vocabulary proposals, which are inadequate for interference-prone tunnel environments. TunnelMIND enhances language-guided defect proposals during inference by utilizing dense visual consistency, converting rough semantic anchors into dependable prompts tailored for specific tunnel hard negatives. The generated masks are then reconstructed into organized defect entities that include details like category, location, geometry, severity, and context. These entities are subsequently linked to retrieval-based explanations and engineering reports, adhering to expert knowledge constraints. The framework processes both visible and ground-penetrating radar (GPR) data, facilitating defect localization, measurement, severity assessment, and documentation without the need for prior training.
Key facts
- TunnelMIND is a training-free framework for tunnel defect inspection.
- It recalibrates language-guided defect proposals using dense visual consistency at inference time.
- The framework reconstructs masks into structured defect entities with category, location, geometry, severity, and context.
- Outputs include retrieval-grounded explanations and engineering-readable reports.
- It operates on visible and GPR data.
- The method addresses interference-heavy tunnel scenes.
- The preprint is available on arXiv with ID 2604.27928.
- The approach supports defect localization, measurement, severity grading, and documentation.
Entities
Institutions
- arXiv