基于局部特征深度信息的绝缘子小样本缺陷检测

Few-shot Insulator Defect Detection Based on Deep Information of Local Features

基于深度学习的目标检测技术已广泛应用于绝缘子缺陷检测中,然而现有目标检测算法主要基于大量缺陷样本训练网络模型,无法对少样本缺陷进行准确识别。针对绝缘子缺陷检测过程中缺陷样本量不足的问题,该文提出了一种基于局部特征深度信息的绝缘子小样本缺陷检测方法。首先采用旋转目标检测网络改进Faster R-CNN(faster region-based convolutional neural network)模型提取绝缘子串区域,然后对绝缘子串特征进行划分,提取绝缘子串局部特征并基于深度推土距离(deep earth mover’s distance,Deep EMD)网络实现小样本缺陷检测。实验结果表明,在玻璃绝缘子自爆缺陷检测中,所提出方法采用2张训练样本可取得与现有目标检测方法 200张训练样本相同的效果,采用10张训练样本的绝缘子自爆检测在与真值框的交并比阈值为0.5至0.95之间的平均精度(mean average precision,mAP)达到0.65,该方法为小样本电力设备缺陷智能化检测提供了新的方法和思路。

Object detection technology based on deep learning has been widely used in the insulator defect detection.However,the existing object detection algorithms are mainly based on the abundant defect samples to train the network models,which is unable to identify the defects with few samples accurately.To solve the problem of insufficient defect samples in the insulator defect detection,this paper proposes a novel few-shot insulator defect detection based on the deep information of local features.Firstly,the insulator strings are extracted using the oriented R-CNN(oriented region-based convolutional neural network).Next,the insulator string features are divided into sub-blocks and the local features are employed to realize the few-shot defect detection based on a deep EMD(earth mover’s distance)network.The experimental results show that the proposed method with 2 training samples can achieve the same results as those of the existing object detection method with 200 training samples for the self-explosion defect detection of glass insulators.The mAP(mean average precision)of insulator self-explosion detection with 10 training samples is up to 0.65.The proposed few-shot defect detection method provides a new solution and an implementation method for the intelligent defect detection of the power equipment with few defect samples.