Automated Multi-Class Structural Damage Detection and Segmentation for Unmanned Aerial Vehicle-Based Infrastructure Inspection Using Deep Learning

Summary

The paper presents a three-stage deep learning pipeline that classifies, segments, and simulates UAV-based inspections of structural damages such as cracks, corrosion, and spalling on civil infrastructure.

Why it matters

Accurate multi-class damage detection is crucial for timely maintenance of bridges and offshore platforms, reducing manual inspection costs. By integrating classification, pixel-level segmentation, and a UAV flight simulation, the framework offers an end-to-end solution capable of near real-time monitoring even on low hardware.

Highlights

• Stage 1 fine-tunes ResNet50 with varying learning rates and class weights, achieving 85.69% accuracy on a test set of 1,600 images.
• Stage 2 employs DeepLabV3+ with a partially frozen backbone, reaching an mIoU of 0.7890 for three-class segmentation.
• Stage 3 simulates a boustrophedon UAV inspection, detecting 16 out of 20 damaged regions and delivering an average inference time of 16.29 ms per frame.

Critical take

The study relies heavily on publicly available datasets and only 315 real UAV images, which may not fully capture the variability of field conditions such as lighting changes and occlusions.