1. What is the complete detection workflow?

  • Capture: The drone flies along pre-planned routes to capture full-bridge high-definition photos with a 100MP medium-format telephoto camera. A 500m-class bridge generates nearly 100GB of data and over ten thousand photos.
  • Upload: After the flight, all photos are uploaded to the platform for unified processing.
  • AI screening: The AI model automatically screens surface defects (cracks, honeycomb, pitting, exposed rebar) and filters them by defect type, size and severity level.
  • Manual review: Professional engineers review the AI results; only verified defects enter the report, ensuring reliable determination.
  • 3D mapping & report: Verified defects are mapped onto the 3D bridge model by component and output as standardized inspection reports.
This model ensures the detection model can be centrally iterated and updated, producing more stable recognition quality.

2. Why is detection performed on the platform instead of on the drone?

A single 100MP photo reaches 40-50MB. With such a data volume, edge-side compute and storage on the drone cannot support real-time AI detection. Centralized platform detection also allows the model to be iterated and updated, keeping recognition quality stable.
Method
Capability
Applicable to
Edge real-time detection
Limited, data volume unsupported
Lightweight scenarios
Platform unified detection
Iterable model, stable accuracy
Standard bridge inspection workflow

3. How is detection accuracy ensured?

  • Imaging capability: The 100MP medium-format telephoto camera captures 0.1mm-level crack details at engineering standard from a safe distance; 0.05mm width has been verified with a caliper at 3 meters.
  • Training data: The model is trained on more than 200,000 crack defect samples; common defects such as honeycomb, pitting and exposed rebar are even easier to recognize with higher accuracy.
  • Human-in-the-loop: AI completes the initial screening; professional engineers perform the final review, combining automatic efficiency with professional judgment.
  • Multi-period comparison: The same defect is tracked across inspection rounds on the 3D model, supporting change observation and treatment priority.

4. Applicable conditions and scenarios

  • Full-bridge automated screening of surface defects
  • Workflow separating field capture from office-based detection, relying on platform compute and a unified AI model
  • Inspection projects with large data volumes and high accuracy requirements
  • Routine inspections requiring multi-period defect comparison

5. Notes

AI identification targets surface defects; conclusions on structural defects must be comprehensively determined by professional engineers in accordance with codes. Detection accuracy depends on imaging quality, shooting distance and model coverage; special defect types may require model iteration or custom training, and defect rating should follow platform unified detection plus manual review.