基于EfficientNet-YOLOv5s的绝缘子缺陷检测

Insulator Defect Detection Based on EfficientNet-YOLOv5s Network

针对目前复杂背景下绝缘子缺陷小目标检测准确率低的问题,提出一种深度学习框架下的EfficientNet-YOLOv5s神经网络检测算法,首先通过无人机航拍输电线路中含有各类绝缘子的图像,并通过图像增强技术丰富图像数据集,然后用EfficientNet网络替换YOLOv5s主干网络,用改进的网络对标注的绝缘子数据集进行训练和测试,最后对模型的损失函数和非极大值抑制算法加以改进,进一步解决绝缘子目标重叠导致的漏检问题。实验结果表明,改进的网络平均精度达到98.5%,满足输电线路中绝缘子缺陷检测要求。

Aiming at the problem of low detection accuracy of small target of insulator defect in complex background at present,an EfficientNet-YOLOv5s neural network detection algorithm based on deep learning framework is proposed.Firstly,the images of various insulators in the transmission line are aerial taken by UAV,and the image data set is enriched by image enhancement technology.Then,the EfficientNet network is used to replace the YOLOv5s backbone network.The improved network is used to train and test the labeled insulator data set.Finally,the non-maximum suppression algorithm and loss function of the model are improved to further solve the problem of missing detection caused by overlapping insulator targets.The experimental results show that the mAP of the improved network reaches 98.5%,which meets the requirements of insulator defect detection in transmission lines.