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

作　　者：陶攀 方宇 王欣 杨梅[1] 闵帆[1] 胡玲 Tao Pan;Fang Yu;Wang Xin;Yang Mei;Min Fan;Hu Ling(School of Computer Science and Software Engineering,Southwest Petroleum University,Chengdu,610500,China;School of Computing and Artificial Intelligence,Southwest Jiaotong University,Chengdu,611756,China)

机构地区：[1]西南石油大学计算机与软件学院,成都610500 [2]西南交通大学计算机与人工智能学院,成都611756

出　　处：《南京大学学报(自然科学版)》2024年第5期776-784,共9页Journal of Nanjing University(Natural Science)

基　　金：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(23XNSYSX0084,23XNSYSX0062);西南石油大学2021年一流本科课程培育建设项目(X2021YLKC035)

摘　　要：轨道缺陷检测是保障轨道交通安全运维的关键任务,现有的基于机器视觉的检测方法主要针对轨道图像进行分割,存在模型时间复杂度高、背景噪点干扰大、分割效果不佳等问题.针对以上问题,提出一种改进Segment Anything Model (SAM)的多任务图像分割模型(Multi-Task Advanced SAM,MASAM),有效地提高了训练效率和缺陷分割准确率.首先,通过目标检测模块来确定缺陷目标范围,获取目标边界坐标;其次,将目标边界坐标转换为稀疏嵌入;最后,将稀疏嵌入与SAM模型中Image Encoder模块处理后的图像特征向量输入Mask Decoder,得到缺陷掩膜预测结果 .实验结果表明,在多任务轨道缺陷检测中,MASAM模型的预测效率和准确率均优于其他模型.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 multitask 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 

分 类 号：U216.3[交通运输工程—道路与铁道工程] TP391.41[自动化与计算机技术—计算机应用技术]