南京信息工程大学学报2026，Vol.18Issue(3)：302-309,8.DOI:10.13878/j.cnki.jnuist.20250314002
融合多尺度混合注意力与迁移学习的全卷积网络路面裂缝检测算法
A fully convolutional network integrating multi-scale hybrid attention and transfer learning for pavement crack detection
摘要
Abstract
To address the limitations of conventional crack detection algorithms,such as multi-scale feature loss and high sensitivity to background interference in complex pavement scenarios,this paper proposes a novel pavement crack detection algorithm named HA-FCN-TL.The algorithm is based on a Fully Convolutional Network(FCN)in-tegrated with multi-scale Hybrid Attention(HA)and Transfer Learning(TL).First,an FCN backbone is construc-ted using a pre-trained ResNet34 model,where the transfer learning strategy accelerates model convergence and en-hances feature representation.Second,a hybrid attention module is designed to integrate Convolutional Block Atten-tion Module(CBAM)with self-attention during the encoding stage,achieving a synergistic optimization that en-hances microscopic crack edges while preserving macroscopic topological continuity.This effectively suppresses noise interference from pavement stains,uneven illumination,and other disturbances.Finally,a multi-scale feature fusion mechanism is introduced,employing skip connections to aggregate shallow details and deep semantic informa-tion across layers.Experiments on the DeepCrack dataset demonstrate that the proposed method excels in fractured texture repair and the detection of weak cracks,providing a highly robust solution for pavement structural safety as-sessment in complex environments.关键词
路面裂缝检测/全卷积网络(FCN)/迁移学习/混合注意力/多尺度特征融合/卷积块注意力模块(CBAM)/自注意力Key words
pavement crack detection/fully convolutional network(FCN)/transfer learning(TL)/hybrid attention(HA)/multi-scale feature fusion/convolutional block attention module(CBAM)/self-attention分类
信息技术与安全科学引用本文复制引用
李卓轩,陈彬,杨光,时欣利..融合多尺度混合注意力与迁移学习的全卷积网络路面裂缝检测算法[J].南京信息工程大学学报,2026,18(3):302-309,8.基金项目
国家自然科学基金重点项目(61833005) （61833005）
国家重点研发计划(2020YFA0714300) （2020YFA0714300）
