基于深度卷积神经网络的目标检测算法进展
    点此下载全文
引用本文:张索非,冯烨,吴晓富.基于深度卷积神经网络的目标检测算法进展[J].南京邮电大学学报:自然科学版,2019,39(5):72~80
摘要点击次数: 209
全文下载次数: 122
作者单位
张索非 南京邮电大学 物联网学院江苏南京210003 
冯烨 南京邮电大学 通信与信息工程学院江苏南京210003 
吴晓富 南京邮电大学 通信与信息工程学院江苏南京210003 
基金项目:国家自然科学基金(61701252)和江苏省高校自然科学研究(16KJB510032)资助项目
中文摘要:目标检测是计算视觉领域一个重要的研究方向,主要解决了图像中各种目标的分类和定位等视觉基本问题。由于近年来深度学习特别是深度卷积神经网络研究的兴起,各类基于神经网络的目标检测算法不断地刷新标准检测数据集的最好性能成绩。文中主要分析比较了几种目前主流的基于卷积神经网络的目标检测算法,包括各种两阶段检测器(RCNN,Fast RCNN,Faster RCNN,Mask RCNN,TridentNet)以及单阶段检测器(YOLO,SSD,CornerNet,ExtremeNet)。文中首先介绍了评价目标检测算法常用的数据集以及对应的性能评价指标,然后对上述检测算法的具体实现方式逐一进行分析。最终,在几个检测数据集上我们对不同算法进行实验复现并综合比较了其性能。实验结果表明,主流的检测算法在速度以及性能方面各有侧重点,需要根据实际场景进行选择和权衡。
中文关键词:目标检测  卷积神经网络  两阶段检测器  单阶段检测器
 
Recent advances on object detection using deep CNNs:an overview
Abstract:The object detection is an important research direction in the field of computer vision,it mainly solves the essential questions,such as the classification and location of various targets in images.Due to the rise of deep learning in recent years,especially the deep convolutional neural networks (CNNs),object detection methods based on CNNs keep refreshing the performance records on different benchmark datasets.This paper analyzes several prevail CNN based object detection algorithms,including two-stage detectors,such as RCNN,Fast RCNN,Faster RCNN,Mask RCNN and TridentNet,as well as one-stage detectors,such as YOLO,SSD,CornerNet and ExtremeNet.Firstly,the common benchmark datasets and a corresponding performance evaluation system are proposed.Then,the detailed algorithm and the implementation of aforementioned detection approaches are analyzed.Finally,this paper attempts to reproduce different methods on several detection datasets,and empirically compare their comprehensive performances,such as mean average precision and inference speed.Experimental results demonstrate that these prevail detection methods retain different features on the accuracy and the speed.Thus,it needs to be selected and balanced according to specific application scenarios by a suitable object detection method.
keywords:object detection  convolutional neural network  two-stage detector  one-stage detector
查看全文  查看/发表评论  下载PDF阅读器

你是第2221711访问者
版权所有《南京邮电大学学报(自然科学版)》编辑部
Tel:86-25-85866913 E-mail:xb@njupt.edu.cn
技术支持:本系统由北京勤云科技发展有限公司设计