铁道科学与工程学报2025，Vol.22Issue(1)：429-442,14.DOI:10.19713/j.cnki.43-1423/u.T20240458
改进YOLOv7和SeaFormer的桥梁裂缝识别与测量
Improved YOLOv7 and SeaFormer bridge cracks identification and measurement
摘要
Abstract
In order to address issues such as large model parameters,low detection efficiency,and high latency in previous bridge crack detection models,as well as poor image quality captured by drones,this paper introduced an image quality evaluation model and the SeaFormer lightweight semantic segmentation algorithm.The YOLOv7 object detection algorithm was improved,and an integrated model for bridge crack recognition and segmentation based on the improved YOLOv7 and SeaFormer was trained and established.Additionally,a pixel-level calculation method for crack length and width based on inscribed circles was proposed.Simultaneously,a recognition-segmentation strategy was employed,where cracks were first identified and then the identified crack regions were segmented using a segmentation model,followed by post-processing of the segmented images and mapping back to the original image according to the position information of the target boxes,significantly improving the detection efficiency of cracks.Using the quality assessment model proposed in this paper,publicly available concrete bridge crack images were selected,annotated,and used as the dataset for this model.Through training and testing on this dataset compared with mainstream models,the advantages of the proposed algorithm in terms of accuracy and lightweight were demonstrated.Furthermore,accuracy validation was conducted using images of structural cracks collected by DJI Phantom 4 Pro-v2.0 drones at a distance of 3 meters from the surface to be tested.The relative error in crack width detection is within 18%,and within 10%for crack length detection.In addition,in a mixed test set including images of pillar cracks from Zhongshan Bridge in Lanzhou City,the crack recognition localization accuracy reaches 91.38%,F1 score is 88.94%,and recall rate is 86.62%.The crack segmentation accuracy is 93.66%,with an intersection over union of 90.17%.The research results indicate that the integrated bridge crack detection method based on improved YOLOv7 and SeaFormer achieves a smaller model size and higher detection efficiency while maintaining both speed and accuracy,making it more suitable for crack detection in structures such as bridges and towers using mobile devices like drones.关键词
桥梁工程/裂缝检测/无人机/深度学习/轻型多层感知模块/YOLOv7模型/SeaFormer模型Key words
bridge engineering/crack detection/unmanned aerial vehicle/deep learning/lightweight multi-layer perception/YOLOv7 model/SeaFormer model分类
交通工程引用本文复制引用
杨国俊,齐亚辉,杜永峰,石秀名..改进YOLOv7和SeaFormer的桥梁裂缝识别与测量[J].铁道科学与工程学报,2025,22(1):429-442,14.基金项目
国家自然科学基金资助项目(51808274,52168042) （51808274,52168042）
甘肃省科技计划资助项目(22JR5RA250) （22JR5RA250）
