基于改进YOLOV4网络的绝缘子缺陷检测

Inspection Method for Insulator Defect Based on Improved YOLOV4 Network

针对现有目标检测方法进行绝缘子缺陷检测速度慢、精度低等问题,改进YOLOV4网络提高绝缘子缺陷检测性能。对比主干特征提取网络不同输出层添加卷积注意力模块(CBAM)缺陷检测结果以确定注意力机制引入方法;采用K⁃Means聚类算法确定适合绝缘子特征的锚框尺寸;加强特征提取网络采用CSPlayer并利用深度可分离卷积代替普通卷积,减少模型参数,提高检测速度;同时,加强特征提取网络中添加金字塔池化模块(SPP),融合多重感受野缺陷特征,改善检测精度;增大类别损失比重,提高分类精度;采用柔性非极大值抑制代替普通非极大值抑制,避免小目标缺陷重叠引起漏检。实验结果表明,改进YOLOV4网络的平均精度均值mAP和检测时间分别为92.26%和19.82 ms,满足绝缘子缺陷检测精度和速度要求。

The improved YOLOV4 network is presented in order to overcome the problems of low speed and low accuracy of existing ob⁃ject detection methods for insulator defect inspection.The attention mechanism introducing method is determined by comparing defect inspection results of different convolutional lack attention modules(CBAM)added in the output layers of backbone network.Moreover,the size of anchor suitable for insulator features is determined by using K⁃Means clustering algorithm.CSPlayer is used and ordinary convolution is replaced by separated convolution in the strengthen feature extraction network to reduce model parameters and increase inspection speed.Furthermore,pyramid pooling module(SPP)is added in the feature extraction network to integrate defect features with multiple receptive fields and improve detection accuracy.The classification accuracy is improved by increasing the proportion of category loss.In addition,ordinary non⁃maximum value suppression is replaced by the flexible one to avoid missed detection caused by overlap⁃ping of small defects.Experimental results demonstrate that the mean average precision(mAP)and detection time of improved YOLOV4 are 92.26%and 19.82 ms respectively,meeting accuracy and speed requirements for insulator defect inspection.