面向电力巡检的轻量化目标检测与故障识别方法研究

Research on Lightweight Object Detection and Fault Recognition Method for Power Inspection

目前电力巡检的2个基本需求是硬件资源占用率低和实时检测速率高,然而现有技术仍面临计算成本与检测速率的挑战,因此基于YOLO系列目标检测算法,设计了可在轻量化平台上部署的网络模型Power Net。Power Net中设计了轻量化特征提取模块Shuffle CSP和特征图感知融合模块Fuse,取得了检测速率与精度的良好折中。最终的模型可达到76.8%的平均准确率,且参数量和浮点运算次数仅有5.5 M和12.8 G,在TESLA V10016 GB GPU和Intel i5-8400 CPU上的检测速度分别达到200~400 fps和2~5 fps,符合实时检测的要求。

At present,the two basic requirements of power inspection are low hardware resource occupancy and high real-time detection rate.However,the existing technology still faces the challenges of computing cost and detection rate.Therefore,based on Yolo series object detection algorithms,a network model Power Net which can be deployed on lightweight platform is designed.A lightweight feature extraction module Shuffle CSP and a feature map sensing fusion module named Fuse are designed in Power Net,and a good compromise between detection rate and accuracy is achieved.The final model has an average accuracy of 76.8%,and the number of parameters and floating-point operations are only 5.5 M and 12.8 G.the detection speed on Tesla V10016 GB GPU and Intel i5-8400 CPU reaches 200~400 fps and 2~5 fps respectively,which meets the requirements of real-time detection.