基于深度卷积神经网络的高速铁路积雪深度判识方法

Identification method of snow depth along high-speed railway based on deep convolutional neural network

针对高速铁路线路上积雪深度的动态判识问题,提出一种基于铁路综合视频图像识别的积雪深度判识方法.首先,对通过综合视频监控系统获得的雪深图像进行处理,利用U-Net神经网络进行图像分割,建立轨道上的雪深数据集.然后,对雪深数据集进行标注,将雪深图像分为100 mm以下、100 mm~轨面和高于轨面3个类别.在此基础上提出基于DenseNet-201深度卷积神经网络模型的雪深图像识别方法.最后,对模型进行验证.研究结果表明:对于光线较好的图像,采用DenseNet-201深度卷积神经网络模型的识别准确率达到93.57%.相较于VGG-16、ResNet-50等模型识别结果,虽然DenseNet-201深度卷积神经网络模型计算耗时长于ResNet-50模型,但是,识别准确率较ResNet-50、VGG-16模型分别提高了2.08%和4.24%.研究成果可为高速铁路沿线积雪深度的动态掌握提供技术支撑.

To address the issue of dynamic snow depth recognition on high-speed railway tracks,this paper proposes a snow depth identification method based on comprehensive railway video image recog-nition.Firstly,the snow depth images obtained from the comprehensive video monitoring system are processed.The U-Net neural network is used to segment the images,thereby establishing a dataset of snow depths on the tracks.Subsequently,the snow depth dataset is annotated by categorizing the snow depth images into three classes:below 100 mm,100 mm to the rail surface,and above the rail surface.Based on this dataset,a snow depth image recognition method is established using the DenseNet-201 deep convolutional neural network model.Finally,the model is validated.The re-search results indicate that for images with good lighting conditions,the recognition accuracy of the DenseNet-201 deep convolutional neural network model reaches 93.57%.Compared to the recogni-tion results of other models like VGG-16 and ResNet-50,although the DenseNet-201 deep convolu-tional neural network model has a longer computation time than the ResNet-50 model,it improves rec-ognition accuracy by 2.08%and 4.24%compared to ResNet-50 and VGG-16 models,respectively.The research results can provide technical support for the dynamic identification of snow depth along high-speed railways.