基于改进YOLOX的轻量级钢轨表面缺陷检测算法

Lightweight Rail Surface Defect Detection Algorithm Based on Improved YOLOX

针对现有基于深度学习的钢轨表面缺陷检测方法在嵌入式检测系统上兼容性较差、计算资源占用高以及检测速度慢的问题,提出了一种基于改进YOLOX的轻量级钢轨表面缺陷检测算法。模型中主干特征层以MobileNetv3单元为基础,在保留其网络轻量化的同时进行局部优化,改进了浅层网络的激活函数,嵌入了SE(Squeeze and Excitation)注意力机制;在加强特征层优化了尾部的冗余卷积。通过与几种代表性算法进行对比试验,验证该算法的性能。结果表明:本文提出的改进算法在模型参数量仅为1.10×106的情况下,检出率和准确率分别达到了92.17%和90.92%,每秒传输帧数(Frame Per Second,FPS)为115.07,模型大小仅为原模型的1/5。该算法在保证较高检测精度的同时大大降低了模型参数量,并提升了检测速度,更适合部署于算力有限的嵌入式轨道检测系统,可为钢轨缺陷高效检测提供有效手段。

A lightweight rail surface defect detection algorithm based on improved YOLOX was proposed to address the issues of poor compatibility,high computational resource consumption,and slow detection speed of existing deep learning based rail surface defect detection methods on embedded detection systems.In the model,the backbone feature layer is based on the MobileNetv3 unit.While retaining its network lightweight,it carries out local optimization,improves the activation function of the shallow network,and embeds the SE(Squeeze and Exception)attention mechanism.The redundant convolution at the tail was optimized by strengthening the feature layer.The performance of this algorithm was verified by conducting comparative experiments with several representative algorithms.The results show that the improved algorithm proposed in this article has a model parameter quantity of only 1.10×106,the detection rate and accuracy reached 92.17%and 90.92%,respectively.The Frame Per Second(FPS)was 115.07,and the model size was only 1/5 of the original model.This algorithm greatly reduces the number of model parameters while ensuring high detection accuracy,and improves detection speed.It is more suitable for deployment in embedded track detection systems with limited computing power,which can provide an effective means for efficient detection of rail defects.