基于Faster RCNN的桥梁缺陷检测研究  

Research on Bridge Defects Detection Based on Faster RCNN

作　　者：杨洋 张华 YANG Yang;ZHANG Hua(School of Intelligent Manufacturing and Information,Jiangsu Shipping College,Nantong 226010,China;Education Informatization Management Center,Jiangsu Shipping College,Nantong 226010,China)

机构地区：[1]江苏航运职业技术学院智能制造与信息学院,江苏南通226010 [2]江苏航运职业技术学院教育信息化管理中心,江苏南通226010

出　　处：《江苏航运职业技术学院学报》2024年第3期38-43,共6页Journal of Jiangsu Shipping College

基　　金：江苏航运职业技术学院科技项目(HYKY/2023B04)。

摘　　要：尽管近年来目标检测技术已取得显著进展,但在复杂环境中的多目标检测仍面临诸多挑战。针对Faster RCNN模型在桥梁检测中遇到的问题,提出三点改进:通过采用ResNet101作为特征提取网络,取代传统的VGG16,以此缓解因网络深度增加而导致的信息传递衰减问题,提高特征学习的效率;通过引入递归特征金字塔结构,更有效地处理不同尺度的目标,从而增强检测性能;通过在模型中嵌入注意力机制,进一步强化模型对关键区域的识别能力并减少背景噪声的影响,使其能够更加聚焦于目标特征。经过改进,模型的准确率提升至92.5%,平均精度达到91.5%。Although target detection technology has made significant progress in recent years,multi-target detection in complex environments still faces many challenges.To address the problems encountered by the Faster RCNN model in bridge detection,three aspects are proposed for improvement:by adopting ResNet101 as the feature extraction network instead of the traditional VGG16,it is to alleviate the problem of attenuation of information transfer due to the increase in the depth of the network,and to improve the efficiency of feature learning;by introducing a recursive feature pyramid structure,different scales of targets can be dealt with more efficiently,thus to enhance the detection performance;by embedding the attention mechanism in the model,it further strengthens the model’s ability to recognize key regions and reduces the influence of background noise,so that it can focus more on target features.As a result of the improvements,the accuracy of the model was increased to 92.5%,with an average accuracy of 91.5%.

关 键 词：桥梁缺陷 目标检测 Faster RCNN 

分 类 号：U446[建筑科学—桥梁与隧道工程]