基于改进训练策略的YOLOv10接触网绝缘子缺陷检测模型优化方法

An Improved Training Strategy Based YOLOv10 Method for Catenary Insulator Defect Detection

摘要: 针对接触网绝缘子缺陷检测任务中模型训练稳定性、效率与精度难以兼顾的问题，本文提出一种基于改进训练策略的YOLOv10检测方法，通过对YOLOv10n模型进行结构简化与两阶段冻结微调优化，系统研究了不同训练策略对缺陷检测性能的影响。实验结果表明，两阶段冻结微调策略能有效提升模型检测精度，在自建绝缘子缺陷数据集上，mAP@0.5达到0.717，相比基线模型提升2.14个百分点；而过度简化模型结构则会导致性能显著下降，mAP@0.5相比基线降低20.09个百分点。可见本研究提出的冻结微调改进策略，在不改变模型源码的情况下实现了性能提升，为绝缘子缺陷检测的工程化应用提供了有效技术方案，并以期能为基于YOLOv10乃至其他YOLO版本的类似工业缺陷检测提供一定的训练策略参考。

Abstract: To address the challenge of balancing training stability, efficiency and accuracy in catenary insulator defect detection, this research proposes an improved training strategy based on YOLOv10. Through structural simplification and two-stage frozen fine-tuning optimization of the YOLOv10n model, we systematically investigated the impact of different training strategies on detection performance. Experimental results show that the two-stage frozen fine-tuning strategy effectively improves model detection accuracy, achieving mAP50 of 0.717 on a self-built insulator defect dataset, which is 2.14 percentage points higher than the baseline model. Meanwhile, excessive simplification of the model struc- ture leads to significant performance degradation, with mAP50 dropping by 20.09 percentage points compared to the baseline. Thus, the research provides an effective training strategy for YOLOv10-based insulator defect detection, achieving performance improvement without changing the original model code, providing an effective solution for such application, while highly expected to offer some fine-tuning strategy references for similar industrial defect detection based on YOLOv10 and other YOLO versions.