Lightweight crack detection in complex environments via model pruning and knowledge distillation

Cracks result in safety risks to sewage treatment structures. Although existing deep learning-based detection models have achieved high accuracy, their complexity presents significant challenges for deployment on mobile devices. To improve automation, accuracy, and efficiency of crack detection in complex environments, a lightweight model named YOM3_PD is proposed by integrating YOLOv5, MobileNetV3 architecture, model pruning, and knowledge distillation. Experimental results show that, after structural pruning and knowledge distillation, the model parameters and size are decreased by 87.4 % and 86.4 %, respectively, while maintaining high detection accuracy under low-light and high-noise conditions. Compared to conventional models, YOM3_PD delivers higher accuracy, smaller size, and faster inference, making it highly efficient and practical for real-time deployment on mobile devices.