基于深度学习的图像目标检测方法具有检测精度高、检测速度快等优点,广泛应用于路面病害检测中,目前研究多关注俯视影像的路面病害检测,前视影像中复杂场景对检测精度影响的研究尚不足.本文基于YOLOv8(you only look once version 8)提出一种路面病害检测模型 YOLO-RMID(road maintenance inspection detection).利用Mask掩模将天空与地面分隔开,屏蔽空中悬挂输电线区域;将注意力机制融入主干特征提取部分中的快速空间金字塔池化(spatial pyramid pooling fast,SPPF)模块,提高裂缝所在区域权重;在特征融合部分中通过将BiFusion模块与RepBlock模块相结合,构建多尺度融合特征BFRB(BiFusion RepBlock)结构,提高模型对路面病害的感知能力;为验证方法可行性,制作路面病害数据集LNTU_RMID,结合公开数据集RDD2022,与常用的MUENet、CrackYOLO及DGE-YOLO-P模型进行对比评价.结果表明,本方法的综合性能相对最优,平均精度分别提高了约6.7％、5.4％、6.6％.
A method for detection of pavement defects in front-view images based on improved YOLOv8
In highway maintenance and inspection,the precise defects is a critical task that the precise detection of pavement defects is a critical task that directly impacts road safety and longevity.However,the complexity of scenes captured in road front-view images during maintenance inspections often results in low accuracy for pavement disease detection.This issue not only impedes the effectiveness of maintenance work but also poses potential risks to road users.To address this crucial challenge,this paper presents a pavement disease detection model called YOLO road maintenance detection(YOLO-RMID),which builds upon the YOLOv8 model to overcome the limitations of existing methods.One significant obstacle faced by the model was distinguishing between hanging wires and cracks in the images.This confusion can lead to inaccurate detection and misclassification.To resolve this,a well-designed mask module was employed to separate the sky from the ground,effectively isolating these two elements and reducing the likelihood of mistaking power lines for cracks.This method significantly enhances crack detection accuracy and aids in distinguishing different image objects.The model might incorrectly identify damage to pavement markings and crack repair errors as actual cracks,leading to false positives and inaccurate assessments of pavement conditions.To address this,the ASPPF module is ingeniously designed(noting that spatial pyramid pools operate swiftly).By focusing on specific features and patterns,the module enables the model to better differentiate true cracks from other anomalies,thereby reducing false detections and providing more reliable results.For partially or fully shaded pavement diseases,the model often struggles to identify them accurately.Shadows can obscure vital features,making it difficult for models to detect and classify diseases.To mitigate this,a BiFusion RepBlock(BFRB)structure is proposed to lessen the impact of shadows by enhancing the model's capability to handle occluded areas.This structure facilitates feature extraction even in shadowed regions,improving pavement disease detection accuracy.To thoroughly assess the YOLO-RMID model's performance,a comprehensive dataset named LNTU-RMID is prepared from original images collected by the highway maintenance inspection system.In the LNTU_RMID dataset,the YOLO-RMID model is rigorously compared with prevalent models used in highway pavement inspection:MUENet,CrackYOLO,and DGE-YOLO-P.The overall performance of the YOLO-RMID detection model is markedly superior to these current models.Specifically,the mean average precision(mAP)is increased by approximately 6.7％,5.4％,and 6.6％respectively compared to these models.This significant performance enhancement underscores the efficacy and superiority of the YOLO-RMID model in managing complex scenes and accurately detecting pavement defects.
NETL
NSTL
万方数据
