基于深度学习与WAFF特征融合的桥梁损坏研究  

Research on bridge damage detection on deep learning and WAFF feature fusion

作　　者：张幽迪 吕卫东[1] 郝月华 段雪艳 ZHANG Youdi;LV Weidong;HAO Yuehua;DUAN Xueyan(School of Mathematics and Physics,Lanzhou Jiaotong University,Lanzhou 730070,China)

机构地区：[1]兰州交通大学数理学院,兰州730070

出　　处：《计算机应用文摘》2025年第4期67-70,74,共5页

基　　金：国家自然科学基金:多自由度非光滑脉冲耦合振子的锁频共振与参数振动:分析与计算(11962011)。

摘　　要：为了解决传统人工检查桥梁损坏类型存在的主观性和效率低的问题,文章提出了一种基于深度学习的Res-Swin-WAFF模型,用于桥梁损坏的细粒度分类。该模型结合了Swin Transformer网络和改进的ResNet18网络,分别用于提取图像的全局特征和局部特征。通过融合全局和局部特征,模型能够有效地利用不同层次的信息,从而更好地识别物体之间的细微差别。同时,文章引入了位置注意力模块(PAM),使模型能够关注到缺陷的具体位置,提升了模型的理解能力和局部细节感知能力。此外,文章设计了一个WAFF特征融合模块,该模块将全局和局部特征进行融合,实现了特征的高效利用。实验结果表明,与传统的单一模型相比,所提WAFF特征融合模块和位置注意力模块(PAM)双支路Res-Swin-WAFF模型在桥梁分类任务中表现更为优秀。该模型能够综合考虑全局信息,并关注局部细节,使其位置特征信息更为丰富,空间理解能力更强,分类性能也得到了显著提升。In order to address the subjectivity and low efficiency of traditional manual inspection of bridge damage types,this article proposes a Res-Swin-WAFF model based on deep learning for fine-grained classification of bridge damage.This model combines Swin Transformer network and improved ResNet18 network,which are used to extract global and local features of images,respectively.By integrating global and local features,the model can effectively utilize information at different levels to better identify subtle differences between objects.At the same time,the article introduces the Position Attention Module(PAM),which enables the model to focus on the specific location of defects,thereby improving the model̓s understanding ability and local detail perception ability.In addition,the article designs a WAFF feature fusion module that integrates global and local features,thereby achieving efficient utilization of features.The experimental results show that the proposed dual branch Res-Swing-WAFF model of WAFF feature fusion module and position attention module(PAM)performs better in bridge classification tasks than the traditional single model.This model can comprehensively consider global information and pay attention to local details,with richer location feature information,stronger spatial understanding ability,and significantly improved classification performance.

关 键 词：缺陷分类及检测 Resnet Swin Transformer 位置注意力 WAFF特征融合 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]