融入注意力的YOLOv3绝缘子串识别方法

YOLOv3 Identification Method Incorporating Attention for Insulator String

绝缘子串是输电线路系统中至关重要的器件之一,基于图像处理的绝缘子串识别,能够更快更准确地识别到绝缘子串。针对输电线路中绝缘子串与绝缘子串、绝缘子串与背景之间相互干扰导致目标识别时存在误差较大的问题,提出一种融入注意力机制的YOLOv3绝缘子串识别方法。该方法在YOLOv3模型的基础上,首先在特征提取网络Darknet⁃53中加入无参注意力SimAM,聚焦网络和增强有效特征,抑制干扰特征,提高网络对绝缘子串的注意能力;其次根据绝缘子串的形状特征,调整模型预设Anchor box的值;最后以焦点损失(Focal loss)作为置信度和分类损失的损失函数,解决了正负样本分布不均衡的问题。实验结果表明,该方法解决了绝缘子串与绝缘子串、绝缘子串与背景之间相互干扰导致识别不准确的问题,识别精度达到了97.89%,模型具有较好的识别性能。

Insulator string is one of the most important elements in the transmission line system.Identification of in⁃sulator string based on image processing can identify insulator string quickly and accurately.In view of large error in the target identification which is caused by the mutual interference between insulator strings and between insulator string and background in the transmission line,a method of YOLOv3 insulator string identification with attention mechanism is proposed.Based on the YOLOv3 model,first,the non⁃parametric attention SimAM is added into the feature extraction network Darknet⁃53 to focus on the network,enhance the effective features,suppress the interfer⁃ence features and improve the attention ability of the network to the insulator string.Then,the preset value of Anchor box of the model is adjusted in accordance with the shape characteristics of insulator string.Finally,the focal loss is used as the loss function of confidence and classification loss to solve the problem of unbalanced distribution of posi⁃tive and negative samples.The experimental results show that this method solves the problem of inaccurate identifica⁃tion caused by mutual interference between insulator strings and between insulator string and background.The iden⁃tification accuracy reaches 97.89%and the model has good identification performance。