RRecherchily
← Back to projects
ActiveMarch 2026

Computer Vision for Structural Health Monitoring in Civil Engineering: Automatic Crack Detection and Quantification via Deep Segmentation Networks

Manual visual inspection of concrete infrastructure (bridges, buildings, dams) is costly, subjective, and hazardous. This project develops an automatic crack detection and quantification system from drone-captured images, using semantic segmentation architectures (U-Net, DeepLabV3+) trained on an annotated dataset of Algerian infrastructure.

Computer visionCivil engineeringSemantic segmentationDrone inspection

Team

AM
Abdelhamid MASSERITI
Research Director
NR
Nadia RAHMANI
Researcher

Methodology

  1. Collect and annotate a dataset of 3,000 crack images from concrete structures in Algeria (bridges, public buildings) with pixel-level segmentation.
  2. Train and compare three architectures: U-Net (ResNet-50 backbone), DeepLabV3+ (MobileNetV2), and SegFormer-B2, with adapted data augmentation (rotations, lighting variations).
  3. Develop a post-processing module for geometric crack quantification (width, length, orientation, relative area).
  4. Integrate into a drone data processing pipeline (DJI Mavic 3) with georeferencing of detected defects.

Evaluation metrics

  • Mean Intersection over Union (mIoU) on crack class — target > 78%.
  • Crack width measurement accuracy — average error < 0.5 mm compared to manual measurement.
  • Inference time per image (1920x1080) on embedded GPU (Jetson Orin Nano) — target < 200 ms.

Budget

$1,200 for drone rental and image capture campaigns, $500 in GPU credits for training, $300 for annotation (supervised crowdsourcing).

Expected deliverables

  • Annotated DZ-CrackSeg dataset published on Hugging Face (first crack dataset on Algerian infrastructure).
  • Web application for inspection result visualization with interactive map.
  • Paper submitted to ECCV 2027 Workshop on Civil Infrastructure.
← Back to projects