基于改进YOLOv5-LITE轻量级的配电组件缺陷识别

Defect Identification of Distribution Components Based on Improved YOLOv5-LITE Lightweight

为对配电组件缺陷进行精确快速的定位和识别,提出一种基于改进YOLOv5-LITE轻量级的配电组件缺陷识别方法。为使模型便于部署至移动设备终端,该方法使用ShuffleNetV2作为骨干网提取特征构建YOLOv5-LITE轻量级神经网络模型,并摘除ShuffleNetV2的1024卷积和5×5池化,采用全局平均池化操作替代,降低网络参数量,提升模型检测速度;通过引入有利于细粒度目标检测的152×152特征层,实现了对大、中、小尺度的缺陷检测;在PANet架构中采用深度可分离卷积代替下采样使得网络更加轻量化。实验结果表明:该方法能够识别电缆脱离垫片、电缆与绝缘子脱落、无环绝缘子3种缺陷,其检测精度分别达到92%、95%、95%,网络参数量约为YOLOv5的1/4,检测速度达到2 ms/张。所提出的方法具有实时性、准确率高、轻量化等特点。

In order to accurately and quickly locate and identify the defects of distribution components,a lightweight defect identification method of distribution components based on improved YOLOv5-LITE is proposed.To make the model easy to deploy to mobile device terminals,this method uses Shufflenetv2 as the backbone network to extract features,constructs YOLOv5-LITE lightweight neural network model,and removes 1024 convolution and 5×5 Pooling of Shufflenetv2,which is replaced by global average pooling operation to reduce the amount of network parameters and improve the speed of model detection.By introducing the 152×152 feature layer,which is conducive to the detection of fine-grained objects,the defect detection of large-,medium-and small-scales is realized.Using deep separable convolution instead of downsampling in PANet architecture makes the network more lightweight.The experimental results show that this method can be adopted to identify three defects:cable separation gasket,cable and insulator falling off and acyclic insulator.The detection accuracy is 92%,95%,and 95%,respectively.The amount of network parameters is about 1/4 of YOLOv5,and the detection speed is 2 ms/piece.The proposed method has the characteristics of real-time,high accuracy and light weight.