基于改进CycleGAN与YOLOv8s的混凝土坝水下裂缝识别方法  

Underwater Crack Identification for Concrete Dams Based on Improved CycleGAN and YOLOv8s

作　　者：赵阳[1] 康飞[1] 万刚[2] ZHAO Yang;KANG Fei;WAN Gang(School of Infrastructure Engineering,Dalian University of Technology,Dalian 116024,China;China Yangtze Power Co.,Ltd.,Yichang 443000,China)

机构地区：[1]大连理工大学建设工程学院,辽宁大连116024 [2]中国长江电力股份有限公司,湖北宜昌443000

出　　处：《水电能源科学》2025年第4期158-162,共5页Water Resources and Power

基　　金：国家重点研发计划(2022YFB4703404);广西重点研发计划(AB24010003);国家自然科学基金项目(52079022)。

摘　　要：针对受水下环境影响造成的混凝土坝水下裂缝图像获取困难、样本稀缺,裂缝检测效率低、精度差、主观性强等问题,提出基于生成对抗网络CycleGAN和目标检测网络YOLOv8s的水下裂缝检测方法。首先,引入梯度惩罚WGAN-GP损失与相似性度量LPIPS损失,提出一种改进的CycleGAN图像风格迁移网络,以此生成高质量水下裂缝图像,解决数据样本不足的问题;之后,添加SimAM无参注意力并引入WIoU损失,提出改进的YOLOv8s水下裂缝识别网络,以提高水下裂缝图像识别的精度。试验结果表明,改进CycleGAN方法起到了良好的数据扩充作用,能有效提升后续检测任务的精度;改进YOLOv8s方法在消融、对比试验中,裂缝识别精度较原网络、Faster R-CNN、YOLOX-s、YOLOv5s分别提高2.4%、5.4%、2.4%、1.2%,检测效果满足高效、精确的要求,可为混凝土坝水下裂缝识别提供技术支持。Aiming at the problems of difficult image acquisition,scarcity of image samples,low efficiency,poor accuracy and subjectivity of crack detection for underwater cracks in concrete dams caused by the influence of the underwater environment,this study proposes an underwater crack detection method based on generative adversarial network CycleGAN and objtect detection network you only look once(YOLOv8s).Firstly,an improved CycleGAN image style migration network is proposed by introducing the Wasserstein Generative Adversarial Networks-Gradient Penalty(WGANGP)loss and the Learned Perceptual Image Patch Similarity(LPIPS)loss as a means to generate high-quality underwater fracture images,and solve the problem of insufficient data samples;After that,an underwater crack recognition network based on improved YOLOv8s is proposed to improve the accuracy of underwater crack image recognition by adding the parameter-free attention mechanism(Simple Attention Mechanism,SimAM)and introducing the cross-entropy monotonic focusing mechanism Wise Intersection over Union(WIoU)loss;The experimental results show that the improved CycleGAN method plays a good role in data expansion,and can effectively improve the accuracy of the subsequent detection task;The improved YOLOv8s method improved the crack identification accuracy by 2.4%,5.4%,2.4%and 1.2%compared with the original network,Faster R-CNN,YOLOX-s and YOLOv5s in the ablation and comparison experiments,and the detection effect can satisfy the requirements of high efficiency and accuracy,which can provide technical support for the identification of underwater cracks in concrete dams.

关 键 词：水下裂缝识别 生成对抗网络 数据扩充 损失函数 注意力机制 

分 类 号：TV698.1[水利工程—水利水电工程]