摘要
目的 针对传统隧道衬砌裂缝检测效率低、成本高、周期长以及当前基于计算机视觉、图像处理技术的裂缝智能检测方法效率低、精度低、检测结果不准确等问题,提出一种改进的网络模型YOLO v5-IBX对公路隧道衬砌裂缝进行智能检测。方法 在原始YOLO v5网络模型的检测层中新增一个低维尺度和在特征提取层中融入注意力机制,提高特征融合利用和对小目标的检测精度,降低网络参数的计算量,达到减少裂缝细节信息丢失的目的;对采集到的公路隧道衬砌裂缝图像,通过图像翻转、裁剪、调整图像饱和度、对比度等随机转换方式来进行数据增强,增加数据特征样本,建立数据集,以满足模型检测的需求;在建立的隧道衬砌裂缝数据集上进行试验,以精确率、召回率、计算平均精度及平均精度均值作为检测精度的综合评价指标,将笔者提出的网络模型YOLO v5-IBX与原始的YOLO v5等其他网络模型进行对比。结果 采用改进的网络模型YOLO v5-IBX检测隧道衬砌裂缝,在迭代300次的情况下,训练损失可以降到0.014,裂缝检测精度率达到97.8%左右,召回率达到97.7%左右,精度均值达到98.6%左右,均优于其他模型,检测精度得到有效提高。结论 相比较传统的人工检测方法和原始YOLO v5检测算法,改进的网络模型YOLO v5-IBX可以更快速、准确地识别出隧道衬砌裂缝,为隧道衬砌裂缝检测提供新的更加实用的检测方案。
Aiming at the problems such as low efficiency,high cost and long cycle of traditional tunnel lining crack detection,low efficiency,low precision and inaccurate detection results of current intelligent crack detection methods based on computer vision and image processing technology,this study proposed an improved network model YOLO v5-IBX for intelligent crack detection of highway tunnel lining.A new low-dimensional scale is added to the detection layer of the original YOLO v5 network model and attention mechanism is integrated into the feature extraction layer,which improves the feature fusion utilization and detection accuracy of small targets,reduces the calculation amount of network parameters,and achieves the purpose of reducing the loss of crack details.The collected highway tunnel lining crack images are enhanced by random conversion methods such as image flipping,clipping,adjusting image saturation,contrast,etc.to increase data feature samples and build data sets to meet the needs of model detection.The network model YOLO V5-IBX proposed in this study was compared with other network models such as the original YOLO v5,using the accuracy rate,recall rate,calculated average accuracy and average accuracy mean as comprehensive evaluation indexes of detection accuracy.The test results show that the network model YOLO v5-IBX proposed in this study can detect tunnel lining cracks.In the case of 300 iterations,the training loss can be reduced to 0.014,the crack detection accuracy rate is about 97.8%,the recall rate is about 97.7%,and the average accuracy is about 98.6%,all of which are better than other models.The detection accuracy is effectively improved.Compared with the traditional manual detection method and the original YOLO v5 detection algorithm,the network model YOLO V5-IBX proposed in this study can identify tunnel lining cracks more quickly and accurately,providing a new and more practical detection scheme for tunnel lining crack detection.
作者
何兆益
常宝霞
吴逸飞
李冬雪
HE Zhaoyi;CHANG Baoxia;WU Yifei;LI Dongxue(School of Transportation,Chongqing Jiaotong University,Chongqing,China,400074;School of Information Engineering,Shaanxi Fashion Engineering University,Xi′an,China,712046;Chongqing Fengjian Expressway Co.Ltd.,Chongqing,China,404600)
出处
《沈阳建筑大学学报（自然科学版）》
CAS
北大核心
2023年第5期888-898,共11页
Journal of Shenyang Jianzhu University：Natural Science
基金
国家重点研发计划项目(2018YFB1600201)
交通运输部科技示范项目(2021-581)。
