融合多级特征与注意力机制的路面裂缝检测  

Pavement crack detection fuses multi-level features and attention mechanism

作　　者：苏天成 郑津津[1] 张广强 丰穗 张健康 周洪军[2] SU Tiancheng;ZHENG Jinjin;ZHANG Guangqiang;FENG Sui;ZHANG Jiankang;ZHOU Hongjun(Department of Precision Machinery Instrumentations,University of Science and Technology of China,Hefei 230026,China;National Synchrotron Radiation Laboratory,Hefei 230029,China)

机构地区：[1]中国科学技术大学精密机械与精密仪器系,安徽合肥230026 [2]国家同步辐射实验室,安徽合肥230029

出　　处：《传感器与微系统》2024年第6期157-160,共4页Transducer and Microsystem Technologies

基　　金：国家自然科学基金联合基金资助项目(GG2090090072,U1332130,U1713206);国家自然基金重大仪器专项项目(61727809);安徽省重点研究与开发计划资助项目(1704a0902051);国家重点研发项目(SQ2019YFC010463)。

摘　　要：针对深度学习模型应用在道路裂缝检测时,存在裂缝提取不完整及检测速度慢等问题,提出了一种基于ResNet34骨干网络并结合通道注意力和空间注意力机制对特征图进行多级特征融合学习的算法。提出的算法由特征提取网络、多级特征融合模块构成,能够生成清晰准确的裂缝分割图像。其中,特征提取网络提取三原色(RGB)图像的分层级特征,多级特征融合模块学习ResNet34分层级特征信息,且各层的输出采用分层监督方式引导网络快速训练。为证明网络的有效性,在公开裂缝数据集上进行了测试,测试结果显示提出的算法在F1、平均交并比(MIoU)和帧率(FPS)上均超过了其他经典网络。When deep learning model is applied to road crack detection,there are issues including incomplete crack extraction and sluggish detection speed,an algorithm is proposed for multi-level feature fusion learning of feature maps based on the ResNet34 backbone network and combined with the channel attention and spatial attention mechanisms.The algorithm can produce clean and precise crack segmentation images since it is made up of a feature extraction network and a multi-level feature fusion module.The feature extraction network extracts the hierarchical features of RGB images,the multi-level feature fusion module learns the hierarchical feature information of ResNet34,and the output of each layer is supervised in a hierarchical manner to guide the rapid training of the network.To verify the effectiveness of this network,tests are carried on the public crack dataset,test results show that the algorithm surpasses exiting classical methods in F1,MIoU and FPS.

关 键 词：多级特征融合 注意力机制 裂缝检测 图像分割 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TN911.73[自动化与计算机技术—计算机科学与技术]