- 5damage classesEfflorescence, rebar exposure, spalling, break and crack
- 5architectures benchmarkedSegFormer, YOLOv8, U-Net, DeepLabV3 and EfficientNet
- Pixel-levelmasksShape and extent of each defect — not just its presence
The challenge
Structural inspection of bridges, piers and walls relies on people reviewing images by eye. The damage types look nothing alike — a hairline crack is a few pixels wide, spalling covers large irregular regions, efflorescence is a pale stain — and the scanner platform needed one model that handles all of them.
Our approach
We framed the task as multi-class semantic segmentation, so every pixel is assigned a damage class and the output can be measured, not just flagged. Five architectures — transformer, single-stage detector and encoder–decoder families — were trained and compared on the client’s imagery to find the best trade-off for the scanner.
Workflow
How it works, step by step.
The pipeline in 5 stages, from capture to inspect. Each stage has a clear input, a clear output and a reason to exist.
- 01
Capture
Surface imagery of concrete structures from the GPR & vision scanner.
- 02
Label
Pixel-level annotation of five damage classes, including thin structures like cracks.
- 03
Benchmark
SegFormer, YOLOv8, U-Net, DeepLabV3 and EfficientNet trained and compared on the same data.
- 04
Segment
The selected model outputs a color-coded mask per damage class for every image.
- 05
Inspect
Overlays show inspectors where damage is and how far it extends, ready for reporting.
Results
What it produced.
- Rebar exposure
- Spalling
- Efflorescence
- Crack
Stack & techniques
- Semantic segmentation
- SegFormer
- YOLOv8
- U-Net
- DeepLabV3
- EfficientNet
- Python
Delivered by members of the Saerosense founding team. Client names are withheld unless public — references are available on request.
Related service
Start a project
Have a similar problem?
Send a few lines about your data and the outcome you want. We reply within two business days with a feasibility read and a proposed scoping plan.