中国铁道科学2025，Vol.46Issue(2)：37-48,12.DOI:10.3969/j.issn.1001-4632.2025.02.04
基于动态蛇形卷积的钢轨RGB图像光带分割方法
Segmentation Method of Wheel-Rail Contact Surface for RGB Rail Images Based on Dynamic Snake Convolution
摘要
Abstract
To address the issue that jagged edges of the wheel-rail contact surface lead to low accuracy of segmentation algorithm for grayscale rail images,this paper proposes a segmentation method for wheel-rail contact surface in RGB rail images based on dynamic snake convolution.A rail image acquisition module based on a color linear array camera and a white laser light source is designed and implemented.By embedding dynamic snake convolution,this research enhances the extraction of jagged irregular features and improves the DeepLabv3+segmentation network,thereby completing integrated segmentation of wheel-rail contact surface and rail surface based on RGB image,as well as classification and detection of the wheel-rail contact surface.The experimental results show that the average Intersection over Union(IoU)for RGB rail images excluding turnout section segmentation is 93.5%,the average pixel accuracy of the category is 96.39%,and the pixel accuracy is 98.85%.For RGB images containing the turnout section segmentationt,the average IoU,the average pixel accuracy of the category and the pixel accuracy of the are 91.87%,96.04%,and 98.60%,respectively.RGB images can better represent the true state of the wheel-rail contact surface.The segmentation network improved by adding dynamic snake convolution can enhance the accurate extraction of the rail-wheel rail contact surface area,the average IoU is improved by 2.25%compared to existing methods.关键词
钢轨/光带分割/检测/RGB图像/动态蛇形卷积Key words
Rail/Segmentation of wheel-rail contact surface/Detection/RGB image/Dynamic snake convolution分类
交通运输引用本文复制引用
程雨,刘金朝,姜昕良,张长伦,张国粹,顾子晨,王乐,宋浩然..基于动态蛇形卷积的钢轨RGB图像光带分割方法[J].中国铁道科学,2025,46(2):37-48,12.基金项目
中国铁道科学研究院集团有限公司院基金课题(2023YJ100) （2023YJ100）
