基于改进YOLO v7的铁路隧道仰坡排水沟病害检测方法

Detection Method for Defects of Drainage Ditch on Railway Tunnel Front Slope Based on Improved YOLO v7 Model

针对铁路隧道洞口仰坡地质条件复杂,排水沟时常堵塞问题,基于改进YOLO v7模型构建了铁路隧道仰坡排水沟状态智能检测方法。将原YOLO v7模型检测头中的传统卷积替换为全维动态卷积,使模型更有效捕捉和识别排水沟图像中的小目标,增强模型对特征的提取能力。将自适应矩阵估计优化器更换为自适应矩阵估计-权重衰减优化器,可解决原模型可能出现的梯度消失问题,显著提高模型的泛化能力。采用层归一化算法替换批量归一化算法,有助于增强模型训练的稳定性。将修正线性单元激活函数调整为高斯误差线性单元激活函数,能够提高模型识别的平均精度。利用无人机拍摄的排水沟图像,将改进后模型与现有常见模型对五类检测目标的识别结果进行了对比。改进后模型对五类目标识别的精确率大部分在0.9以上,且其识别的平均精度、召回率均高于现有常用模型。改进后模型识别精确率更高,检测稳定性更好,满足铁路隧道仰坡排水沟检测需求。

Aiming at the complex geological conditions and frequent blockage of drainage ditches on the front slope of railway tunnel entrance,an intelligent detection method for the status of drainage ditches on the front slope of railway tunnels was constructed based on improved YOLO v7 model.Replacing the traditional convolution in the detection head of the original YOLO v7 model with omni-dimensional dynamic convolution could make the model more effective in capturing and recognizing small targets in drainage ditch images,which enhances the model extraction ability to extract features.The adaptive moment estimation optimizer was replaced with the adaptive moment estimation with weight decay optimizer and could solve the problem of gradient vanishing that may occur in the original model,significantly improving the model generalization ability.The layer normalization algorithms were adopted to replace the batch normalization algorithms,which help to enhance the stability of model training.Adjusting the rectified linear unit activation function to Gaussian error linear unit activation function could improve the average accuracy of model recognition.This paper compared the recognition results of the improved model with existing commonly used models for five types of detection targets using drainage ditch images captured by drones.The accuracy of the improved model for recognizing five types of targets is mostly more than 0.9,and its average recognition accuracy and recall are higher than existing commonly used models.The improved model has higher recognition accuracy and better detection stability,meeting the detection requirements of railway tunnel front slope drainage ditch.