Lightweight YOLO Network for Power Insulator Defect Detection

To address the challenges of high model parameters and computational complexity in existing insulator detection models, which hinder their deployment on mobile inspection devices such as unmanned aerial vehicles with limited computational and storage resources, this paper designs a lightweight YOLOv5s network for detecting self-explosive defects in power insulators. The proposed network reconstructs the architecture using depth-wise separable convolutions, CenterNetinspired backbone design, and an adaptive spatial pyramid module to reduce model parameters and complexity. Structured pruning is then applied to the lightweight model to trim redundant convolutional layers, further compressing the model. Meanwhile, knowledge distillation is employed to compensate for the detection performance degradation caused by pruning. Experimental results on an insulator defect dataset show that the proposed model significantly reduces model complexity and improves inference speed while maintaining detection accuracy. The lightweight architecture achieves a substantial reduction in computational overhead without compromising the ability to accurately identify the defects, demonstrating its suitability for real-time deployment on resource-constrained edge devices.