基于改进YOLOv5s的绝缘子定位检测及红外故障识别

Insulator Positioning Detection and Infrared Fault Recognition Based on Improved YOLOv5s

在绝缘子定位检测和热故障识别中,由于绝缘子红外图像的背景干扰严重,导致平均识别准确率低,为了实现精确定位、检测绝缘子位置和提高识别其热故障的可靠性和准确性,提出了一种基于改进YOLOv5s的绝缘子定位检测和红外故障识别方法。首先,将全局上下文注意力机制与YOLOv5s Backbone部分的C3结构进行融合,提出一种新的结构——C3GC,增强了模型提取特征的能力,并且减少了其计算量。其次,将损失函数替换为VariFocal Loss,提升了模型的召回率,解决了模型漏检的问题。最后,通过引入转置卷积,动态地学习需要补充的参数,减少了模型在上采样过程中特征的丢失,提升了检测效果。实验与测试结果表明,改进后的方法与原YOLOv5s相比,定位精度提升了1.3个百分点,针对故障点的检测精度提升了4个百分点,平均精度提升了2.8个百分点,并且其精确率和召回率均有提升。

In insulator positioning detection and thermal faults recognition,due to severe background interference in insulator infrared images,the average recognition accuracy is low.In order to achieve precise position and detect of insulator position and improve the reliability and accuracy of identifying its thermal faults,an improved insulator positioning detection and infrared faults recognition method based on YOLOv5s is proposed.Firstly,a new structure C3GC is proposed by integrating the global context attention mechanism with the C3 structure of YOLOv5s Backbone,which enhances the ability of the model to extract features and reduces its amount of calculation.Secondly,replacing the loss function with VariFocal Loss,the recall rate of the model is improved,which can reduce the problems of missed detections of model.Finally,by introducing transposed convolution and dynamically learning the parameters that need to be supplemented,the loss of features of the model during sampling process is reduced.The experimental and testing results show that compared with the original YOLOv5s,the improved method improves positioning accuracy by 1.3%,detection accuracy for fault points by 4%,average accuracy by 2.8%,and both accuracy and recall rate is improved.