基于YOLOv8改进算法的电力绝缘子缺陷检测方法

A Power Insulator Defect Detection Method Based on Improved YOLOv8 Algorithm

摘要:

针对无人机巡检时绝缘子缺陷因目标小、种类多、尺度差异大造成的漏检、误检问题,本文提出了一种YOLOv8改进算法,实现在多类型、多尺度绝缘子缺陷检测中准确性的提升；在改进的算法结构中,在回归损失计算中使用WIoU,降低图像质量引起的梯度增益,增强了模型的定位性能和泛化能力；在主干网络的特征提取和预测阶段,引入多尺度混合注意力机制(Multiscale-hybrid Attention,MHA),增强了网络模型学习小目标重要特征的能力；在主干网络末端,使用多尺度深度可分离卷积来增强SPPF模块,形成多尺度空间敏感模块(Multi-Scale-Space Sensitive Module, MS3M),能够充分提取含有上下文特征的多尺度信息；通过实验表明,使用改进的YOLOv8算法,绝缘子缺陷检测的mAP值达到了97.3%,相较于基线YOLOv8提升了6.2%,对多类型多尺度的绝缘子缺陷检测达到了更好的效果,对提升电力巡检业务运维效率具有现实指导意义。

Abstract:

In response to the issues of missed detection and false alarms caused by the small size, diverse types, and large scale differences of insulator defects during UAV inspections, an improved YOLOv8 algorithm is proposed to enhance the accuracy of multi-type, multi-scale insulator defect detection. In the improved structure, WIOU is used in regression loss calculation to reduce gradient amplification caused by image quality, enhancing the model"s localization performance and generalization ability. Adding a multi-scale channel-space attention mechanism into the feature extraction and prediction stages of the backbone network enhances the network model"s ability to learn important features for small targets. The Multi-Scale-Space Sensitive Module (MS3M) is constructed by adding multi-scale depthwise separable convolutions into the SPPF module at the end of the backbone network, which effectively extracts multi-scale information with contextual features. Experimental results show that the improved YOLOv8 algorithm achieves an mAP of 97.3% in insulator defect detection, which is a 6.2% improvement over the baseline YOLOv8. This improved performance in detecting multi-type, multi-scale insulator defects offers practical guidance for enhancing the operational efficiency of power inspection services.