首页|基于改进Segformer的混凝土桥梁表观病害轻量级识别方法

基于改进Segformer的混凝土桥梁表观病害轻量级识别方法

扫码查看
原文链接
  • NETL
  • NSTL
  • 万方数据
现有基于语义分割的混凝土桥梁表观病害识别方法具有模型参数量较大、特征提取不充分、分割精度不高等缺点.针对上述问题,提出基于改进Segformer的混凝土桥梁表观病害轻量级识别方法——Segformer-SP,该方法选取 MiT B0作为编码器,并引入语义融合模块(Semantic Fu-sion Module,SFM)和极化自注意力机制(Polarized Self-Attention,PSA).在 Segformer-SP 中,SFM模块对低级特征和高级特征进行融合,以提高低级特征的语义信息量;同时,PSA自注意力机制模块可以在分割过程中捕捉全局上下文信息,解决病害特征不够充分的问题,从而进一步提高分割效果.试验结果表明:Segformer-SP的平均交并比(mIoU)和平均F1分数(mF1)相较于Seg-former-B0分别提高了 2.41％和1.91％;此外,Segformer-SP的mIoU和mF1均优于大部分主流语义分割算法;Segformer-SP的参数量仅为6.09 × 106,FPS为56.54,更适合应用于终端检测设备.
Lightweight Detection Method of Surface Damage on Concrete Bridges Based on Improved Segformer
The existing methods for detecting surface damage on concrete bridges based on semantic segmentation suffer from several drawbacks,including large model parameters,inadequate feature extraction,and low segmentation accuracy.To address these challenges,this study proposes Segformer-SP,a lightweight surface-damage detection method for concrete bridges based on an enhanced Segformer architecture.Segformer-SP adopts the MiT B0 as the encoder and introduces two novel modules:the semantic fusion module(SFM)and polarized self-attention(PSA).In Segformer-SP,the SFM is employed to fuse low-level features with high-level features,thereby enhancing the semantic information captured by the low-level features.Additionally,PSA captures global contextual information during the segmentation process,effectively addressing the issue of insufficient feature representation for damages and further improving the segmentation accuracy.Experimental results demonstrate that Segformer-SP outperforms Segformer-B0,exhibiting a 2.41％increase in mean intersection over union(mIoU)and a 1.91％increase in mean F1-score(mF1).Moreover,Segformer-SP achieves better performance in terms of mIoU and mF1 than most state-of-the-art semantic segmentation models.Notably,Segformer-SP has a significantly reduced parameter count of only 6.09 X 106,while maintaining a high frame-per-second(FPS)rate of 56.54,making it suitable for deployment on terminal detection equipment.

bridge engineeringbridge damage detectionSegformerlightweightfeature fusion

蒋仕新、唐椿程、杨建喜、李昊、熊元俊、李韧、刘新龙、王笛

展开 >

重庆交通大学信息科学与工程学院,重庆 400074

重庆交通大学交通运输学院,重庆 400074

桥梁工程 桥梁病害识别 Segformer 轻量级 特征融合

国家自然科学基金国家自然科学基金重庆市自然科学基金重庆市自然科学基金重庆市教委科学技术研究计划项目重庆市教委科学技术研究计划项目重庆市教委科学技术研究计划项目

6200306362103068CSTB2022NSCQ-MSX1599cstc2020jcyjmsxmX0047KJQN202000726KJQN202100748KJZD-M202000702

2024

10.19721/j.cnki.1001-7372.2024.02.007

中国公路学报
中国公路学会

中国公路学报

CSTPCD北大核心
影响因子：1.607
ISSN：1001-7372
年,卷(期)：2024.37(2)
  • 31