UAVs Images Based Real-Time Insulator Defect Detection with Transformer Deep Learning
Xinlin Liu, Zhuyi Rao, Yunxiang Zhang, Yun Zheng
Abstract
Insulator defect detection is important for the safety operation of the power grid, which can be inspected via the unmanned aerial vehicles (UAVs) patrolling, demanding high accuracy and real-time capability. In response to this requirement, this paper investigates a real-time end-to-end insulator defect detection algorithm, RT-DETR (Real-Time DEtection TRansformer) with the combination of the model compression method based on the parameter quantization and knowledge distillation. In order to reduce the model parameters and accelerate the detection speed, a lighter backbone and a regularization method for refining the attention computation are applied in the model. Further, a quantization training approach which combines the parameter quantization and self-distillation is used for the model compression. The proposed method is trained and validated on an open-source dataset. Experimental results demonstrate that the average mean average precision (mAP) of the proposed method for the insulator defect detection is 99.5%, and the inference speed is 23ms, meeting the requirements for the UAVs real-time inspection.