基于轻量化网络与增强多尺度特征融合的绝缘子缺陷检测

Insulator Defect Detection Based on Lightweight Network and Enhanced Multi-scale Feature Fusion

随着无人机搭载目标检测算法在输电杆塔绝缘子巡检领域的发展,针对绝缘子缺陷检测速度较低,网络复杂度高且缺陷小目标难以准确检测的问题,提出一种基于轻量化网络与增强多尺度特征融合的YOLOv5-3S-4PH模型进行绝缘子缺陷实时检测。首先将重构的ShuffleNetV2-Stem-SPP(3S)网络作为YOLOv5的主干网络,显著减小了网络的参数量和计算量;其次引入针对小目标的增强多尺度特征融合网络以及4个预测头,来增强网络对绝缘子缺陷的感知能力,并结合Mosaic-9数据增强、CIoU损失函数进一步补偿轻量化导致的检测精度损失;最后将其应用到自制绝缘子数据集进行验证。实验结果表明,该文所提出的模型相对于未改进的YOLOv5,全类平均精度提高了3%,检测速度提高了81.8%,参数量、计算量分别压缩了82.4%、67%。因此,所提出的模型更适合部署在无人机平台上进行绝缘子缺陷的实时监测。

With the development of target detection algorithm embedded in UAV for insulator inspection of transmission towers,a YOLOv5-3S-4PH model based on lightweight network and enhanced multi-scale feature fusion is proposed to detect insulator defects in real time in view of the low detection speed,high network complexity and the difficulty of ac-curate detection of small defect targets.Firstly,the reconstructed ShuffleNetV2-Stem-SPP(3S)network is used as the backbone of YOLOv5,which reduces the amount of network parameters and calculation significantly.Secondly,the en-hanced multi-scale feature fusion network for small targets and four prediction heads(4PH)is added to enhance the network’s perception of insulator defects.Combined with Mosaic-9 data enhancement and CIoU loss function,the loss of detection accuracy caused by lightweight is further compensated.Finally,the YOLOv5-3S-4PH model is applied to the self-made insulator dataset for verification.The experimental results show that mean average precision(mAP)is increased by 3%,the detection speed is increased by 81.8%,and parameters and calculation are decreased by 82.4%and 67%com-pared to original YOLOv5 model.Therefore,the proposed model is more suitable for real-time monitoring of insulator defects deployed on UAV platforms.