摘要
针对现有桥梁路面裂缝分割方法对微小变化的裂缝定位不准,分割效果不佳的问题,提出一种基于双分支特征融合增强网络的小样本桥梁路面裂缝分割方法。该方法以支持分支和查询分支的双分支网络建立基线模型,首先利用预训练的Swin Transformer和ResNet-50网络提取支持分支中桥梁路面裂缝图片的多尺度特征。然后,利用多尺度特征增强注意力促进特征之间的交互,并在交互特征上生成原型集。最后,逐位置计算查询特征与原型集的相似度,并根据最大相似度值逐像素分割出查询图片中的裂缝区域。在自建数据集上进行了大量实验,所提出方法实现了72.04%的mIoU和91.32%的FB-IoU,同时获得了95.23%的Precision、95.08%的Recall和95.02%的F1得分,综合性能优于当前主流的分割模型。
To address the issues of inaccurate localization and poor segmentation performance of existing bridge pavement crack segmentation methods for minor variations,a few-shot bridge pavement crack with dual-branch feature fusion enhancement network is proposed.This method establishes a baseline model using a dual-branch network structure with support and query branches.First,the pretrained Swin Transformer and ResNet-50 networks are used to extract multi-scale features from bridge pavement crack images in support branch.Then,a multi-scale feature enhancement attention is utilized to promote interaction between features,and prototype sets are generated on the interacted features.Finally,the similarity between query features and prototype sets is calculated position by position,and the crack regions in query images are segmented pixel by pixel based on the maximum similarity values.Extensive experiments on a self-built dataset demonstrate that the proposed method achieves 72.04%for mIoU and 91.32%for FB-IoU,respectively,and reaches 95.23%Precision,95.08%Recall and 95.02%F1 score,outperforms current mainstream segmentation models in overall performance.
作者
魏嵩锜
王正振
WEI Songqi;WANG Zhengzhen(College of Architectural Engineering,Shanxi Vocational University of Engineering Science and Technology,Jinzhong,Shanxi 030600,China;School of Civil Engineering,Lanzhou University of Technology,Lanzhou,Gansu 730050,China)
出处
《光电子.激光》
北大核心
2025年第12期1265-1272,共8页
Journal of Optoelectronics·Laser
基金
甘肃省自然科学基金(22JR5RA286)资助项目。
关键词
裂缝分割
桥梁路面
特征增强
无参数度量
crack segmentation
bridge pavement
feature enhancement
non-parameter metric
