摘要
【目的】背景复杂的混凝土结构表面裂缝连续性差、识别率低,基于深度学习的裂缝检测方法存在模型参数量大的问题。【方法】为此,结合U^(2)-Net框架构建了一种聚合多尺度信息的轻量级模型U^(2)-Net_Aggregation,用于复杂背景下的裂缝特征学习。该模型通过增加跳跃连接,使得每个解码层均聚合该层以上所有浅层编码特征,以获得足够的特征细节,提升裂缝分割精度;利用深度可分离卷积(Depthwise Separable Convolution, DSC)对原本的残差模块(ReSidual U-blocks, RSU)进行改进,提出了新的残差模块(RSU-DSC-ECA),来降低聚合多尺度信息时带来的模型复杂度提升的问题,其中的通道注意力机制(Efficient Channel Attention, ECA)可提升模型对裂缝区域的敏感性和对复杂背景的抗干扰能力。【结果】在三组裂缝数据集上进行消融试验,改进后的模型(U^(2)-Net_Aggregation)相较于U^(2)-Net在准确率、交并比、综合评价指标上均有优异的表现。为了验证模型对复杂背景中裂缝的识别能力,利用无人机实地采集的某混凝土结构数据进行试验,其检测效果优于FCN、SegNet、U-Net和U^(2)-Net。【结论】改进后的模型相比U^(2)-Net在召回率、交并比和综合评价指标方面分别提高了4.18%、2.97%和2.03%,可借助无人机影像快速准确地检测出裂缝,为结构裂缝检测提供一种新的方法。
[Objective]In view of the poor continuity and low recognition rate of structural surface cracks with complex background,the crack detection method based on depth learning has the problems of large model parameters.[Methods]This paper constructs a lightweight model U^(2)-Net_Aggregation that aggregates multi-scale information based on the U^(2)-Net framework,which is used to learn fracture characteristics in complex background.By adding jump connections,the model enables each decoding layer to aggregate all shallow coding features above the layer to obtain sufficient feature details and improve the accuracy of crack segmentation;Using depthwise separable convolution(DSC)to improve the ReSidual U-blocks(RSU),a new residual module(RSU-DSC-ECA)is proposed to reduce the problem of increasing model complexity when aggregating multi-scale information.The efficient channel attention(ECA)improves the sensitivity of the model to fracture areas and its anti-interference ability to complex backgrounds.[Results]The ablation experiment was carried out on three sets of fracture data sets.Compared with U^(2)-Net,the improved model(U^(2)-Net_Aggregation)has excellent performance in precision,intersection over union and f1-measure.To verify the model′s ability to identify cracks in complex backgrounds,experiments were conducted using concrete structure data collected by UAV,which outperformed FCN,SegNet,U-Net and U^(2)-Net.[Conclusion]The improved model improved by 4.18%,2.97%and 2.03%in recall,intersection over union and f1-measure,respectively,compared to U^(2)-Net,which can quickly and accurately detect cracks with the help of UAV images,providing a new method for structural crack detection.
作者
程浩东
李怡静
李玥康
胡强
王姣
CHENG Haodong;LI Yijing;LI Yuekang;HU Qiang;WANG Jiao(School of Infrastructure Engineering,Nanchang University,Nanchang 330031,Jiangxi,China;Jiangxi Academy of Water Science and Engineering,Nanchang 330029,Jiangxi,China)
出处
《水利水电技术（中英文）》
北大核心
2024年第6期159-171,共13页
Water Resources and Hydropower Engineering
基金
江西省自然科学基金项目(20232BAB204091)
国家自然科学基金项目(41501454)
江西省水利厅科技项目(202123YBKT25)。
