基于大型卷积核模型和自监督预训练的路面质量分析方法  

Large-scale Convolutional Kernel Model and Self-supervised Pre-training for Road Quality Analysis

作　　者：孙玉龙 钱增志 曾帅康 康伟德 李胤 姜佳岐 朱炳科 SUN Yulong;QIAN Zengzhi;ZENG Shuaikang;KANG Weide;LI Yin;JIANG Jiaqi;ZHU Bingke(China Railway Construction Group Co.,Ltd.,Beijing 100043,China;Faculty of Architecture,Civil and Transportation Engineering,Beijing University of Technology,Beijing 100124,China;CRCC Construction Industrialization Engineering Laboratory,Beijing 100043,China;Institute of Automation,Chinese Academy of Sciences,Beijing 100190,China)

机构地区：[1]中铁建设集团有限公司,北京100043 [2]北京工业大学城市建设学部,北京100124 [3]中国铁建建筑工业化工程实验室,北京100043 [4]中国科学院自动化研究所,北京100190

出　　处：《无线电工程》2023年第3期527-533,共7页Radio Engineering

基　　金：基于计算机视觉的道路质量评估方法研究。

摘　　要：交通行业围绕加快建设交通强国总体目标,努力打造安全便捷、智慧绿色、经济高效的道路网,全面提升公路运行效率和服务水平,实现路网更高质量、更高效率、更好水平的发展。大数据时代,建造企业需要通过智能化、自动化、数据化的技术手段实现降本增效。针对全自动化路面质量评估和分析的实际场景需求,提出了基于大型卷积核模型和自监督预训练的路面质量分析方法,采集了一个大型的路面病害分割数据集。在路面病害识别阶段,提出了基于重参数化大型卷积核的U型网络结构,实现像素级别的高精度路面病害识别,使用图像修补作为模型的自监督预训练代理任务,针对全卷积编码器和全卷积解码器构建了自监督预训练框架,实现了高精度的模型预训练方法,进而实现了高精度的智能路面病害识别,为智能化路面质量评估和质量管理提供理论和决策依据。To accelerate the construction of powerful transportation of China,the transportation industry strives to build a safe,convenient,smart,green,and cost-effective road network,comprehensively improve the efficiency and service level of road operation,and achieve the development of road network with higher quality,efficiency,and better level.In the era of big data,construction enterprises need to reduce costs and increase efficiency through intelligent,automated and data-based technology.A road quality analysis method based on the UNet model with large-scale convolutional kernel and self-supervised pre-training is proposed according to the actual scene requirements of fully automated road quality evaluation and analysis.A large-scale road damage segmentation dataset is collected.In the stage of road disease identification,a UNet model based on a re-parameterized large convolutional kernel is proposed to realize the pixel-level road damage identification,and image repainting is used as the self-supervised pre-training agent task.A self-supervised pre-training framework is built for the fully convolutional encoder and decoder to achieve high-precision model pre-training methods,and in turn realize high-precision intelligent road damage identification.The method provides a theoretical and decision-making basis for intelligent road quality evaluation and quality management.

关 键 词：智能交通 图像分割 U型结构 自监督 质量管理 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]