基于YOLOM算法的路面病害轻量化检测方法  

Lightweight Pavement Disease Detection Based on YOLOM Algorithm

作　　者：胡晓伟[1] 闫奕昕 王大为 张宇辉 HU Xiao-wei;YAN Yi-xin;WANG Da-wei;ZHANG Yu-hui(School of Transportation Science and Engineering,Harbin Institute of Technology,Harbin 150090,Heilongjiang,China;School of Architecture,Harbin Institute of Technology(Shenzhen),Shenzhen 518055,Guangdong,China)

机构地区：[1]哈尔滨工业大学交通科学与工程学院,黑龙江哈尔滨150090 [2]哈尔滨工业大学(深圳)建筑学院,广东深圳518055

出　　处：《中国公路学报》2024年第12期381-391,共11页China Journal of Highway and Transport

基　　金：国家重点研发计划项目(2023YFB2603500);深圳市科技计划项目(KJZD20230923115206014)。

摘　　要：准确且轻量化的路面病害检测能够有效降低路面检测设备硬件需求、加快道路巡检效率,对于公路的智慧养护管理高质量发展具有重要意义。针对现有路面病害检测方法存在的精确度不足问题,提出一种融合状态空间模型的YOLOM路面病害轻量化检测方法。首先,针对路面病害检测任务训练特点,设计多扫描模式的视觉Mamba层,调整数据归一化方法,以更适合训练批次较小的图像特征快速提取;设计了并行计算单元,加快网络计算速度,缩短算法训练时间。其次,以YOLOv9为基础,设计了以SSM为核心机制的Mamba聚合特征提取网络层MELAN和空间金字塔Mamba层SPMELAN。通过MELAN提取图像长距离依赖关系,挖掘病害图像的全局特征。使用SPMELAN融合多尺度感受野信息,增强模型对局部细节和全局语义的表达能力,进而提出融合SSM的YOLOM轻量化算法。最后,在RDD2022数据集上进行对比试验。结果表明:YOLOM的F_(1)分数、mAP50值、mAP50-95值和FLOPs较YOLOv9c均有优化;与主流基线模型检测结果比较,YOLOM算法检测精度和推理速度均达到最优值且模型体积和复杂度最小。该方法具有模型规模小、复杂度低和检测精度高的显著优势,有较强的学习能力和泛化性能,可为道路智慧检测提供助力。Accurate and lightweight pavement distress detection can effectively reduce the hardware requirements of pavement inspection equipment and improve road inspection efficiency.This is crucial for high-quality development of smart road maintenance management.To address the issues of low accuracy in existing pavement distress detection methods,a lightweight road-disease detection method based on YOLOM combined with a state-space model(SSM)is proposed.First,based on the training characteristics of road disease detection tasks,a multiscan visual Mamba layer was designed,and the normalization methods were adjusted to quickly extract image features more suitable for small training batches.In addition,parallel computing units were added to accelerate the network computation speed and reduce the training time of the algorithm.Second,based on YOLOv9,the Mamba efficient layer aggregation network(MELAN)with SSM as the core mechanism and spatial pyramid MELAN(SPMELAN)were designed by integrating the attention hiding mechanism of SSM.The long-distance dependencies of the images were extracted using MELAN,and the global features of the disease images were mined.SPMELAN was used to fuse multiscale receptive field information and enhance both the full-size coverage capability of the receptive field of the high target detector and the model's ability to capture local details and global semantics.Consequently,YOLOM combined with SSM was proposed.Finally,comparison experiments were conducted using RDD2022.The results show that the F_(1)-score,mAP50,mAP50-95,and FLOPs of YOLOM are better than those of YOLOv9c.In comparison with the baseline model checking results,YOLOM achieves the highest detection accuracy and inference speed,with the smallest model size and complexity.YOLOM has significant advantages including low weight,low complexity,high detection accuracy,strong learning,and generalization capabilities,which can assist in intelligent road detection.

关 键 词：路面工程 路面病害检测 深度学习 状态空间模型 注意力机制 

分 类 号：U418.6[交通运输工程—道路与铁道工程]