基于改进YOLOv3的电力设备红外目标检测模型

Infrared Object Detection Model for Power Equipment Based on Improved YOLOv3

红外图像检测技术因具有非接触、快速等优点,被广泛应用于电力设备的监测与诊断中,而对设备快速精确地检测定位是实现自动检测与诊断的前提。与普通目标的可见光图像相比,电力设备的红外图像可能存在背景复杂、对比度低、目标特征相近、长宽比偏大等特征,采用原始的YOLOv3模型难以精确定位到目标。针对此问题,该文对YOLOv3模型进行改进:在其骨干网络中引入跨阶段局部模块;将路径聚合网络融合到原模型的特征金字塔结构中;加入马赛克(Mosaic)数据增强技术和Complete-IoU(CIoU)损失函数。将改进后的模型在四类具有相似波纹外观结构的电力设备红外图像数据集上进行训练测试,每类的检测精度均能达到92%以上。最后,将该文方法的测试结果与其他三个主流目标检测模型进行对比评估。结果表明:不同阈值下,该文提出的改进模型获得的平均精度均值优于Faster R-CNN、SSD和YOLOv3模型。改进后的YOLOv3模型尽管在检测速度上相比原YOLOv3模型有所牺牲,但仍明显高于其他两种模型。对比结果进一步验证了所提模型的有效性。

Infrared image detection technology is widely used in monitoring and diagnosing electrical equipment considering its non-contact and fast advantages.It is generally believed that fast and accurate localization of the equipment is the prerequisite for automatic detection and diagnosis.Compared with visible light images of ordinary objects,the infrared images of power equipment have characteristics of complex background,low contrast,similar object features,and large aspect ratio.Besides,the original YOLOv3 model is difficult to accurately locate the objects of power equipment.In view of the above problems,an improved YOLOv3 model was proposed in this paper:cross stage partial module was introduced into the backbone network;the path aggregation network was integrated into the feature pyramid structure of the original model;in addition,this study also added Mosaic data enhancement technology and CIoU loss function.The improved model was trained and tested on four types of infrared image data sets of power equipment with similar corrugated appearance structures,which showed that the detection accuracy of each type can reach more than 92%.Finally,the results were compared and evaluated with the other three mainstream object detection models.The results show that the mean average precisions of the improved model proposed in this paper were better than Faster R-CNN,SSD and YOLOv3.Although the detection speed of the improved YOLOv3 model is sacrificed compared to the original YOLOv3 model,it is significantly higher than the other two models,further verifying the effectiveness of the proposed model in this paper.