南京大学学报（自然科学版）2024，Vol.60Issue(5)：776-784,9.DOI:10.13232/j.cnki.jnju.2024.05.008
基于改进SAM模型的多任务轨道缺陷检测方法
Multi-task track defect detection method based on improved SAM model
摘要
Abstract
Track defect detection is a critical task for ensuring the safe operation and maintenance of rail transportation.Existing machine vision-based detection methods mainly focus on segmentation of railway images,which have high model time complexity,serious background noise interference and poor segmentation effects.This paper proposes an improved multi-task image segmentation model based on Segment Anything Model(Multi-Task Advanced SAM,MASAM)to effectively improve training efficiency and defect segmentation accuracy.First,the defect scope is determined and the boundary coordinates are obtained through the target detection module;then the boundary coordinates are converted into sparse embeddings;finally,the sparse embeddings and the image feature vectors processed by the Image Encoder module in the SAM model are inputted to the Mask Decoder part to obtain the defect mask prediction result.The experimental results show that the prediction efficiency and accuracy of the MASAM model are superior to other models in multi-task railway defect detection.关键词
轨道缺陷检测/SAM/多任务/图像分割/目标检测Key words
track defect detection/SAM/multi-task/image segmentation/object detection分类
信息技术与安全科学引用本文复制引用
陶攀,方宇,王欣,杨梅,闵帆,胡玲..基于改进SAM模型的多任务轨道缺陷检测方法[J].南京大学学报（自然科学版）,2024,60(5):776-784,9.基金项目
国家自然科学基金(62176221,62276215,62276218,62272398),中央引导地方科技发展专项(2021ZYD0003),2021年第二批产学合作协同育人项目(202102211111),南充市-西南石油大学市校科技战略合作专项资金(23XNSYSX0084,23XN-SYSX0062),西南石油大学2021年一流本科课程培育建设项目(X2021YLKC035) （62176221,62276215,62276218,62272398）
