东南大学学报（自然科学版）2025，Vol.55Issue(5)：1319-1327,9.DOI:10.3969/j.issn.1001-0505.2025.05.012
多尺度单目深度信息辅助的铁路桥梁视觉识别方法
Multi-scale monocular depth information-assisted railway bridge vision inspection method
摘要
Abstract
To improve the accuracy and efficiency of computer vision tasks for infrastructure inspection by deep learning models,a multi-scale monocular depth information-assisted railway bridge vision inspection method was proposed.A multi-task neural network was developed.With the assist of the monocular depth map,two tasks including component recognition and damage segmentation were fulfilled simultaneously.To make full use of the correlation at different scales of the depth map and the two tasks,a multi-scale feature in-teraction strategy was proposed based on the multi-task neural network.The multi-scale multi-modal distilla-tion module,cross-scale feature propagation module,and feature aggregation module were developed to share features and achieve multi-scale feature interaction.A large-scale image dataset was used to validate the pro-posed method.The results show that the mean F1-scores of component recognition and damage segmentation are 93.48%and 85.93%,respectively,which are 1.87%and 4.41%higher than those of the single-task net-work without depth information.Compared with task-specific networks trained separately,floating-point op-erations and inference duration of the multi-task network are reduced by 17.89%and 30.48%,respectively.Therefore,the proposed method can improve the accuracy and efficiency of railway bridge vision inspection.关键词
铁路桥梁/构件识别/病害分割/多任务神经网络/多尺度/单目深度图Key words
railway bridge/component recognition/damage segmentation/multi-task neural network/multi-scale/monocular depth map分类
交通工程引用本文复制引用
杨涵,徐庆凯,章金勇,蒋友,于泓川,舒江鹏,徐声亮..多尺度单目深度信息辅助的铁路桥梁视觉识别方法[J].东南大学学报（自然科学版）,2025,55(5):1319-1327,9.基金项目
国家自然科学基金资助项目(W2412092) （W2412092）
宁波市重点研发计划国际科技合作资助项目(2024H013) （2024H013）
浙江省重点研发计划资助项目(2024C01132). （2024C01132）
