摘要
针对高速铁路和普速铁路的塞钉智能检测问题,提出基于YOLO v5的塞钉检测方法。该方法对训练模型采用He方法初始化,采用微调模式使模型收敛,同时使用dropout正则化方法解决过拟合问题。在训练数据时,对多样式塞钉进行有效分类和样本自动化裁剪,并利用塞钉样本的灰度直方图对数据进行增强;采用非冗余多置信度的方式进行检测,通过先验经验和增加反例来减少误判,并在检测时自动累积塞钉样本。该检测方法已投入现场使用,有效满足了铁路智能巡检的需求。
In view of the intelligent inspection of plug pins for high speed and conventional railways,a plug pin inspection method based on YOLO v5 is proposed.In this method,the training model is initialized by He method,which converges the model by fine-tuning mode,and solves the overfitting problem by dropout regularization method.During data training,multiple-style plug pins are effectively classified and samples are automatically cut,and the grayscale histogram of plug pin samples is used to consolidate the data.Non-redundant multi-confidence coefficient methods are adopted for inspection,and misjudgment is reduced by prior experience and additional counter-examples,and plug pin samples are automatically accumulated during inspection.This inspection method has been applied on site,effectively meeting the needs of intelligent railway patrol inspection.
作者
聂海丽
NIE Haili(Science and Technology Research Institute,China Railway Shanghai Group Co.,Ltd.,Shanghai 200071,China)
出处
《中国铁路》
2023年第4期117-123,共7页
China Railway
关键词
铁路塞钉
YOLO
v5
目标检测
图像识别
智能巡检
railway plug pin
YOLO v5
target inspection
image recognition
intelligent patrol inspection