频谱池化与混洗注意力增强的铁路异物轻量检测
Lightweight Detection of Railway Object Intrusion Based on Spectral Pooling and Shuffled-Convolutional Block Attention Module Enhancement
陈永 1王镇 2张娇娇2
作者信息
- 1. 兰州交通大学电子信息与工程学院,甘肃 兰州 730070;兰州交通大学甘肃省人工智能与图形图像处理工程研究中心,甘肃 兰州 730070
- 2. 兰州交通大学电子信息与工程学院,甘肃 兰州 730070
- 折叠
摘要
针对红外弱光环境下铁路异物侵限检测时存在检测精度低、难以实现轻量化实时检测等问题,提出一种注意力增强的轻量化铁路异物检测方法.首先,采用深度可分离卷积改进Darknet53特征提取网络,轻量化提取红外弱光场景下的铁路异物特征;其次,利用语义引导的红外频谱池化进行特征增强,提升红外图像下采样的特征质量;然后,提出混洗注意力机制(shuffled-convolutional block attention module,shuffled-CBAM),实现对关键红外目标的特征提取与融合,提高网络对红外目标检测的精度;最后,采用无锚框轻量化网络完成铁路异物侵限检测输出,克服锚框检测非极大值抑制操作实时性差的缺点,减小计算量的同时提高检测效率.实验结果表明:所提轻量化模型具有较高检测精度,同较改进前模型尺寸减小 179.01 MB,检测速率提升至 39 帧/s,为YOLOv4方法的3.9倍;相较于对比检测方法,本文所提方法能够快速精确地检测出红外铁路异物.
Abstract
In infrared low-light scenes,railway object intrusion detection faces low detection accuracy,and it is difficult to achieve lightweight real-time detection.Therefore,a lightweight detection method of railway object intrusion based on convolutional block attention module(CBAM)enhancement was proposed.Firstly,the Darknet53 feature extraction network was improved by deep separable convolution to achieve lightweight extraction of railway object intrusion characteristics in infrared low-light scenes.Secondly,semantic-guided infrared spectral pooling was used for feature enhancement to improve the feature quality of infrared image downsampling.Then,a shuffled-CBAM was proposed to achieve feature extraction and fusion of key infrared targets and improve the accuracy of infrared target detection.Finally,the lightweight anchor-free network was used to predict the output result of railway object intrusion,overcoming the deficiency of poor real-time performance due to non-maximum value suppression operation with anchor frame detection,and it reduced calculation load and speeded up the detection efficiency.The experimental results show that the lightweight model has higher detection accuracy,and the size of the model is reduced by 179.01 MB after the improvement.The detection rate is increased to 39 frames/s,which is 3.9 times that of the YOLOv4 method.Compared with other detection methods,the proposed method can detect infrared railway object intrusion quickly and accurately.
关键词
异物检测/红外弱光/混洗注意力/轻量化检测/高速铁路Key words
objection intrusion detection/infrared low-light scene/shuffled-CBAM/lightweight detection/high-speed railway引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据