利用轻量级高分辨率特征的隧道裂缝检测方法
    点此下载全文
引用本文:刘宏伟1,杜晓兵1,徐政超2,3,李伟3,师昕2.利用轻量级高分辨率特征的隧道裂缝检测方法[J].计算技术与自动化,2023,(2):144-150
摘要点击次数: 144
全文下载次数: 0
作者单位
刘宏伟1,杜晓兵1,徐政超2,3,李伟3,师昕2 (1. 陕西交通控股集团有限公司宝鸡分公司陕西 宝鸡 7213992. 西安工程大学 计算机科学学院陕西 西安 7100483. 长安大学 信息工程学院陕西 西安 710064) 
中文摘要:针对现有隧道检测车采集图像成像质量不佳,且洞壁存在大量病害相似物干扰的问题,提出了一种基于轻量化的HR(high resolution)-Net框架的隧道病害检测算法TC(tunnel crack)HR-Net。该算法保留了HR-Net的主干语义特征交换子网框架,移除其他分支子网络以降低模型体积。此外,为弥补“剪枝”操作导致的特征损失,在子网分支末端加入了SE模块,对特征的每一通道进行权重划分,以增强特征抽象水平。通过在与精细化标注隧道病害图像数据集上进行验证,本算法的mIoU指标分别达到80.21%与71.22%,高于其他对比算法,并接近HR-Net的检测结果,但是耗时比后者减少了30%。
中文关键词:隧道裂缝检测  语义分割  隧道检测车图像  轻量级HR-Net  语义特征提取
 
Tunnel Crack Detection Method Using Lightweight High-resolution Features
Abstract:To address the problem of poor image quality of the existing tunnel inspection vehicles and the interference of a large number of disease analogues on the cave wall, a tunnel disease detection algorithm TC(tunnel crack) HR-Net is proposed based on a lightweight HR(high resolution)-Net framework. In addition, to compensate for the loss of features due to the “pruning” operation, an SE module is added at the end of the subnet branch to weight each channel of the feature to enhance the level of feature abstraction. Validated on both normal and finely annotated tunnel disease image datasets, the mIoU metrics of this paper are 80.21% and 71.22% higher than other comparative algorithms, and close to the detection results of HR-Net, but with 30% less time consumption.
keywords:tunnel disease detection  semantic segmentation  tunnel inspection vehicle images  lightweight HR-Net  semantic feature extraction
查看全文   查看/发表评论   下载pdf阅读器