基于YOLO的轨道扣件状态检测

State detection of track fastener based on YOLO

建立了基于YOLO的轨道扣件状态检测模型。为了提高检测速度优化了YOLO神经网络的卷积层并运用反向传播算法对损失函数权重参数进行更新。根据VOC数据集格式创建了样本集并利用该样本集上训练及测试模型。检测结果表明:基于YOLO的扣件状态检测的查准率为94.2%,查全率为85.1%,检测速度达到了60f/s,对环境具有很好的鲁棒性,并且运用该方法与另外两种算法进行对比分析得出该模型检测速度最快。

A state detection method of track fasteners based on YOLO is proposed.In order to improve the detection speed,the convolution layer of YOLO neural network is optimized and back propagation algorithm is used to update weighting parameter of loss function.According to VOC dataset format,sample set is created as well as trained and tested the model on the dataset.The comprehensive test results show that the accuracy of YOLO detection algorithm is 94.2%,the recall rate is 85.1%,and the detection speed is up to 60 f/s,which has good robustness to the environment,using this method and the other two algorithms to compare and analyze.It shows that the model has the fastest detection speed.