复杂环境下路面病害识别模型优化研究

Research on Optimization of Pavement Distress IdentificationModel in Complex Environmen

路面病害自动化检测是道路工程领域的热门研究方向。路面病害因其形态的特殊性与背景的复杂性,导致自动化检测存在精度不够高、泛化性差的问题。针对该现象,基于YOLOv5对主干网络和颈部网络中增加高效的特征提取模块C2F和CBAM,且在检测头网络中增加了微型检测器,形成了优化网络结构;通过车载高清相机自主采集大量具备复杂背景的路面图像,共标记67942张路面图像以供模型训练,并且采用Mosaic算法和MixUP算法对自建路面病害数据集进行了数据增强;在训练时优化损失函数。最后通过设置消融实验和对比实验来探究模网络优化措施对路面病害检测模型的精度影响。研究结果表明,应用C2F、CBAM模块能够有效帮助网络提高特征提取能力,增加微型检测头能够在多尺度下加强模型的检测能力,采用以上3种措施优化后的网络模型在精准率和召回率上提升了11.88%和8.69%,mAP取得了0.719的高分。从病害类别角度而言,模型对纵向裂缝、横向裂缝和坑槽提升了检测精度,尤其是坑槽的检测精度。表明了本文模型在识别路面病害类小尺度的对象时具备优异的检测能力。

Automatic detection of pavement distress is a hot research direction in the field of road engineering.Because of the particularity of pavement distress and the complexity of background,there are some problems in automatic detection,such as low accuracy and poor generalization.Aiming at this phenomenon,this paper adopts YOLOv5 as the baseline model,adds efficient feature extraction modules C2F and CBAM to the backbone network and neck network,and adds micro detectors to the detector head network,forming an optimized network structure.A large number of road images with complex background were collected by car-mounted high-definition camera,and 67942 road images were marked for model training,and the self-built road distress data set was enhanced by Mosaic algorithm and MixUP algorithm.The hyperparameter is optimized and the loss function is optimized during training.Finally,the influence of optimization measures of model network on the accuracy of pavement distress detection model is explored by setting ablation experiments and comparative experiments.The research results show that the application of C2F and CBAM modules can effectively improve the feature extraction ability of the network,and the detection ability of the model can be enhanced at multiple scales by increasing the micro-detector.The accuracy and recall of the network model optimized by the above three measures have increased by 11.88%and 8.69%,and the mAP has achieved a high score of 0.719.From the point of view of distress types,this model improves the detection accuracy of longitudinal cracks,transverse cracks and pits,especially pits.It shows that the model in this paper has excellent detection ability in identifying small-scale objects such as pavement distress.