基于多尺度特征融合网络的铁路工机具目标检测

Object detection of railway tool based on multi-scale feature fusion network

图像目标检测是铁路工机具实现自动清点的关键技术。由于铁路运维工作的特殊性,采集的工机具图像通常存在光照度低、目标尺度差异大以及背景复杂等问题。已有的图像目标检测模型无法在铁路工机具检测中获得满意的效果。本文提出一种基于多尺度特征融合增强网络模型,以深度学习目标检测网络Retinanet为基础,构造了特征融合增强模块,可实现对不同尺度铁路工机具的高精度检测。以真实的铁路工机具图像作为数据集开展了实验研究,结果表明提出的模型较Retinanet具有更好的目标检测效果,mAP从97.85%提高到了98.11%,解决了复杂背景下铁路工机具的精确检测难题,为铁路智能运维奠定了技术基础。

Image object detection is crucial for the automatic counting of railway tools.However,collected images of railway tools have the characteristics of low illumination,huge difference in the scale of different objects and complex backgrounds.Existing image object detection methods cannot detect railway tools efficiently.In this paper,we propose a novel object detection model which is able to enhance the object-detection ability by fusing multi-scale features.Based on the object detection deep learning model Retinanet,we construct a feature fusion enhancement module.By fusing features,our model can efficiently detect railway tools with different scales.Experiments were conducted based on real-world datasets.Results show that our method is more efficient than Retinanet,and the mAP is increased from 97.85%to 98.11%.By the accurate detection of railway tools in complex backgrounds,our approach can lay the technical foundation for intelligent railway operation and maintenance.