基于轻量卷积神经网络的高速铁路隧道表面病害筛选算法

Tunnel Surface Defect Filtering Algorithm Based on Lightweight Convolutional Neural Network for High Speed Railway

随着急剧增加的高速铁路隧道检测需求,基于计算机视觉的高速铁路隧道病害识别和健康检测是国内外的新趋势。然而,高速铁路隧道结构表面图像大多数是无病害图像(占比90%以上),剔除大量无病害图像而只保存有病害图像,可以大幅减少图像存贮量和降低高速海量存贮对硬件要求。为此,提出一种基于深度卷积神经网络的隧道表面病害筛选算法,以推理速度和预测精度均衡的残差神经网络ResNet-18作为主干结构,将深度可分离卷积替代标准卷积搭建适用于实时筛选的轻量模型ResNet-DS(Depthwise Separable),采用权重损失函数、静态离线量化进行模型优化,新算法对神经网络轻量化,实现了海量图像高速识别和剔除。结果表明:改进的轻量模型识别精度高达98.67%,与原模型相比较,筛选速度在GPU(RTX 2060Super 8G)上提升22%(10.86 ms/张),在CPU上提升178%(21.20 ms/张)。该研究为病害快速采集系统提供一种实时筛选算法,更好地满足检测需求。

With the sharp increasing of the requirements for detection of high speed railway tunnel,defect identification and health detection of high speed railway tunnel based on computer vision have become a thriving trend at home and abroad.However,most of the surface images of high-speed railway tunnel structure are defect-free images(accounting for more than 90%).Eliminating a large number of defect-free images and saving only defect images can greatly decrease the image storage and reduce hardware requirements because of high-speed massive storage.To solve this problem,an algorithm based on deep convolution neural network is proposed to filter tunnel surface defects.ResNet-18,a residual neural network with balanced inference speed and accuracy,is used as the lightweight backbone architecture.A lightweight model called ResNet-DS is built by replacing standard convolution with depthwise separable convolution for real-time filtering.Weight loss function and static offline quantization are employed to optimize the model.The new algorithm lightens the weight of neural network,realizing high-speed recognition and elimination of massive images.The results show that the accuracy of the improved lightweight model is as high as 98.67%.Compared with the original model,the filtering speed on GPU(RTX 2060 super 8 G)is increased by 22%(10.86 ms per picture)and on CPU by 178%(21.20 ms per picture).This study provides a real-time filtering algorithm for the rapid acquisition system of defects,which can better meet the detection requirements.