Lightweight CNN for Real-Time UAV Bridge Crack Detection
A new lightweight convolutional neural network framework enables real-time crack classification on UAV bridge inspections at 825 FPS with only 11.21M parameters. The model addresses four key challenges: weak crack features, degraded imaging, class imbalance, and limited computational resources. It integrates a lightweight backbone, CBAM attention, inspection-scene augmentation, and Focal Loss. Tests on the SDNET2018 dataset confirm high efficiency.
Key facts
- Proposed framework achieves 825 FPS inference speed
- Model has 11.21 million parameters
- Addresses weak crack features, degraded imaging, class imbalance, and limited compute
- Uses CBAM for channel and spatial enhancement
- Employs directed robust augmentation based on inspection-scene priors
- Uses Focal Loss for hard-sample learning
- Tested on SDNET2018 bridge deck dataset
- Published as arXiv:2604.27617v1
Entities
Institutions
- arXiv