摘要
针对目前传统铁路异物侵线检测算法识别精度不高、对于小尺度目标异物存在漏检等问题,提出一种基于改进Faster RCNN的小尺度铁路侵限算法。在特征提取网络中利用特征金字塔模型将高层特征与低层特征相融合;通过修改锚点框尺寸和增加锚点个数来提高对目标建议区域的精确性;提出一种基于衰减得分的NMS算法;在引入迁移学习思想同时利用在线难例挖掘训练网络以解决数据缺乏、训练难收敛的问题。实验结果表明,改进的Faster RCNN与传统的Faster RCNN网络相比,mAP(mean average precision)提高了2.1%,对小目标的识别有较好准确度。
Aiming at the problems of low recognition accuracy and missing detection of foreign objects in small-scale targets,a small-scale railway intrusion detection algorithm based on improved Fast RCNN was proposed.In the feature extraction network,the feature pyramid model was used to fuse the high-level features with the low-level features.The accuracy of the proposed area was improved by modifying the anchor box size and increasing the number of anchor points.A NMS algorithm based on attenuation score was proposed.The migration learning idea was introduced,and the online difficult case mining training network was used to solve the problems of data shortage and training convergence.Experimental results show that the mean average precision of the improved Fast RCNN is 2.1%higher than that of the traditional Fast RCNN.It has good accuracy for small target recognition.
作者
余志强
张明
YU Zhi-qiang;ZHANG Ming(School of Electrical and Electronic Engineering,Shijiazhuang Railway University,Shijiazhuang 050000,China)
出处
《计算机工程与设计》
北大核心
2022年第4期1023-1031,共9页
Computer Engineering and Design
基金
国家自然科学基金项目(11872257、51674169)
河北省自然科学基金面上基金项目(E2018210144)
河北省重点研发计划基金项目(20354501D)
石家庄市军民融合基金项目(201060104A)。
关键词
铁路异物侵限
小目标检测
特征融合
在线难例挖掘
迁移学习
railway foreign matter intrusion
small target detection
feature fusion
online hard example mining
transfer lear-ning
