基于CutMix数据增强与多约束损失函数的YOLOv7盾构隧道渗漏水检测

YOLOv7 shield tunnel water leakage detection method based on CutMix data augmentation and multi-constraint loss function

盾构隧道强度投影图像中渗漏水尺寸不一致且像素占比偏小,现有目标检测模型的关键特征学习能力较弱,存在渗漏水病害目标检测精度偏低的问题。本文提出了基于CutMix数据增强与多约束损失函数的改进YOLOv7盾构隧道渗漏水检测方法。首先采用镶嵌CutMix方法对隧道图像进行数据增强,将多张不同的训练样本进行随机裁剪,拼接融合成具有综合特征的新样本;然后以YOLOv7网络为骨架结构,引入高效通道注意力模块,提高渗漏水关键特征的自主学习与表达能力;最后引入多约束几何条件的损失函数,提高预测框几何形状的精度,从而提升模型预测精度。在光线良好、光线不佳和存在遮挡等复杂环境情况下,选取Fast R-CNN、SSD、YOLOv5、YOLOv7这4种算法进行对比,试验表明,本文算法渗漏水检测精度达85.90%,平均精度比同类算法分别提高5.55%、8.89%、3.93%、2.75%,具有较高的稳健性和泛化能力。

Since the size of leakage water in the intensity projection image of the shield tunnel is inconsistent and the proportion of pixels is limited,the learning ability of key features of object detection models is weak.As a result,the detection accuracy of leakage disease targets is too low to meet the requirements of the application.Therefore,an improved YOLOv7 shield tunnel leakage water detection method based on CutMix data enhancement and multi-constraint loss function is proposed to address the issue in this paper.Firstly,the tunnel images are enhanced using the embedded CutMix approach.Various training samples are randomly combined into new samples with comprehensive features.Secondly,the YOLOv7 network is employed as the skeletal structure,and an efficient channel attention module is introduced to enhance the ability of crucial leakage features to learn and express themselves autonomously.Finally,a loss function incorporating multi-constraint geometric conditions is designed to improve the accuracy of the geometric shape of the prediction box,thereby improving the model's predictive accuracy.The four algorithms included Fast R-CNN,SSD,YOLOv5,and YOLOv7 are chosen for comparison in complex environments with good lighting,poor lighting,and occlusion.The experiments show that our algorithm achieves a leakage detection accuracy of 85.90%.The average accuracy of the proposed method is higher than Fast R-CNN,SSD,YOLOv5,and YOLOv7 by 5.55%,8.89%,3.93%,and 2.75%,respectively.It exhibits good robustness and generalization ability.