摘要
传统目标检测算法大多基于滑动窗口和人工特征提取,存在计算复杂度高和在复杂场景下鲁棒性差的缺点。近年来,研究人员将深度学习技术应用于目标检测领域,显著提高了算法性能。相比传统算法,基于深度学习的目标检测算法具有速度快、准确性高和在复杂条件下鲁棒性强的优点。从评价指标、公开数据集、传统算法框架等方面对目标检测任务进行阐述,按照是否存在显式的区域建议和是否定义先验锚框两种分类标准,对现有基于深度学习的目标检测算法进行分类,分别介绍算法的演进路线并总结算法机制、优势、局限性及适用场景。在此基础上,分析对比代表性算法在公开数据集中的表现,并对基于深度学习的目标检测的未来研究方向进行展望。
Most existing conventional object detection algorithms are based on sliding windows and artificial feature extraction,and exhibit disadvantages such as high computational complexity and unsatisfactory robustness under complex conditions. Recently,deep learning has been applied to object detection,bringing significant improvements to algorithm performance. Compared with conventional target detection algorithms,deep-learning-based algorithms offer high speed,accuracy and robustness under complex conditions.In this paper,we first expound upon target detection tasks in terms of their evaluation indicators,public datasets,and traditional algorithm frameworks. Then the existing deep learning-based target detection algorithms are categorized based on two criteria,whether there is an explicit region proposal and whether to define a priori anchorbox. We introduce the evolution of these algorithms,summarizing their mechanism,advantages,limits and application scenarios. On this basis,the performance of the representative algorithmson public datasets are analyzed and compared. Finally,we discuss the future directionsofresearch in deeplearning-based object detection.
作者
李柯泉
陈燕
刘佳晨
牟向伟
LI Kequan;CHEN Yan;LIU Jiachen;MU Xiangwei(School of Maritime Economics and Management,Dalian Maritime University,Dalian,Liaoning 116026,China;School of Computer and Information Engineering,Hebei Finance University,Baoding,Hebei 071051,China)
出处
《计算机工程》
CAS
CSCD
北大核心
2022年第7期1-12,共12页
Computer Engineering
基金
教育部人文社会科学项目(18YJC630124)
中央高校基本科研业务费专项资金(3132020241)
辽宁省教育厅科技研究项目(L2014203)
辽宁省社会科学规划基金(L14BGL012)
大连海事大学教学改革项目(2020Y60)。
关键词
目标检测
深度学习
卷积神经网络
计算机视觉
特征提取
object detection
deep learning
convolutional neural network
computer vision
feature extraction