摘要
为提高水下结构物裂缝检测的准确性,基于UNet提出一种改进的检测算法。该方法在编码器中集成ResNet50以强化特征提取,并在解码器后端引入通道-空间并行注意力机制,增强模型对裂缝区域的识别能力。同时,采用Dice loss与BCE loss组合损失函数,优化模型对裂缝区域的敏感度。实验显示,该方法在自建及CFD、DeepCrack数据集上的精确率、F_(1)分数和平均交并比分别达到91.46%,91.04%,84.45%,优于现有主流分割算法。该方法可有效提升水下裂缝检测的效率和准确性,对水下结构健康监测具有一定实际应用价值。
To enhance the accuracy of underwater structural crack detection,an improved detection algorithm based on UNet was proposed.In the method,Res Net50 was integrated into the encoder to strengthen feature extraction and a channel-spatial parallel attention mechanism was introduced in the decoder to enhance the model's recognition of crack areas.Additionally,a combination of Dice loss and BCE loss was used to optimize the model's sensitivity to crack areas.Experiments demonstrated that the method achieved the precision,F_(1)score,and mean Intersection over Union of 91.46%,91.04%,and 84.45%,respectively,on self-built,CFD,and Deep Crack datasets,out performing existing mainstream segmentation algorithms.This method caneffectively improve the efficiency and accuracy of underwater crack detection,and has practical application value for the health monitoring of underwater structures.
作者
贺峰
柏超
张文伟
程风雯
叶云凌
HE Feng;BAI Chao;ZHANG Wenwei;CHENG Fengwen;YE Yunling(CCCC Second Highway Consultants Co.,Ltd.,Wuhan 430052,China;School of Naval Architecture,Ocean and Energy Power Engineering,Wuhan University of Technology,Wuhan 430063,China)
出处
《交通科技》
2025年第6期58-63,共6页
Transportation Science & Technology
关键词
UNet
水下结构检测
裂缝分割
注意力机制
平均交并比
UNet
underwater structure detection
crack segmentation
attention mechanism
mean intersection over union
