电子学报 ›› 2021, Vol. 49 ›› Issue (7): 1428-1438.DOI: 10.12263/DZXB.20200570

• 综述评论 • 上一篇    下一篇

基于深度学习的通用目标检测研究综述

程旭1, 宋晨1, 史金钢2, 周琳3, 张毅锋3, 郑钰辉1   

  1. 1.南京信息工程大学计算机与软件学院,江苏 南京 210044
    2.西安交通大学软件学院, 陕西 西安 710049
    3.东南大学信息科学与工程学院, 江苏 南京 210096
  • 收稿日期:2020-06-15 修回日期:2021-01-15 出版日期:2021-07-25 发布日期:2021-08-11
  • 作者简介:程 旭(通信作者) 男,1983年出生,山西祁县人,南京信息工程大学计算机与软件学院副教授、硕士生导师,研究方向为计算机视觉、图像理解.E‑mail:xcheng@nuist.edu.cn
    宋 晨 男,1997年出生,江苏盐城人,南京信息工程大学计算机与软件学院软件工程专业硕士研究生,主要研究方向为深度学习、目标检测及其应用.
    周 琳 女,1976年出生,江苏镇江人,东南大学信息科学与工程学院副教授、硕士生导师,主要研究方向为深度学习、语音信号处理及其应用.E‑mail:Linzhou@seu.edu.cn
    张毅锋 男,1963年出生,安徽芜湖人,东南大学信息科学与工程学院副教授、硕士生导师,研究方向为计算机视觉、图像理解.E‑mail:yfz@seu.edu.cn
    郑钰辉 男,1982年出生,山西芮城人,南京信息工程大学计算机与软件学院教授、博士生导师,研究方向为计算机视觉、模式识别.E‑mail:zhengyh@vip.126.com
  • 基金资助:
    国家自然科学基金(61802058);江苏省自然科学基金(BK20201267);中国博士后科学基金(2019M651650)

A Survey of Generic Object Detection Methods Based on Deep Learning

Xu CHENG1, Chen SONG1, Jin-gang SHI2, Lin ZHOU3, Yi-feng ZHANG3, Yu-hui ZHENG1   

  1. 1.School of Computer and Software,Nanjing University of Information Science and Technology,Nanjing,Jiangsu 210044,China
    2.School of Software Engineering,Xi’an Jiaotong University,Xi’an,Shaanxi 710049,China
    3.School of Information Science and Engineering,Southeast University,Nanjing,Jiangsu 210096,China
  • Received:2020-06-15 Revised:2021-01-15 Online:2021-07-25 Published:2021-08-11

摘要:

目标检测是计算机视觉领域中最基础且最重要的任务之一,是行为识别与人机交互等高层视觉任务的基础.随着深度学习技术的发展,目标检测模型的准确率和效率得到了大幅提升.与传统的目标检测算法相比,深度学习利用强大的分层特征提取和学习能力使得目标检测算法性能取得了突破性进展.与此同时,大规模数据集的出现及显卡计算能力的极大提高也促成了这一领域的蓬勃发展.本文对基于深度学习的目标检测现有研究成果进行了详细综述.首先回顾传统目标检测算法及其存在的问题,其次总结深度学习下区域提案和单阶段基准检测模型.之后从特征图、上下文模型、边框优化、区域提案、类别不平衡处理、训练策略、弱监督学习和无监督学习这八个角度分类总结当前主流的目标检测模型,最后对目标检测算法中待解决的问题和未来研究方向做出展望.

关键词: 计算机视觉, 深度学习, 目标检测, 卷积神经网络

Abstract:

Object detection is one of the most fundamental and important tasks in the field of computer vision, which is the basis of high?level vision tasks such as behavior recognition and human?computer interaction. With the development of deep learning technology, the accuracy and efficiency of object detectors have been greatly improved. Compared with traditional object detection algorithms, deep learning utilizes powerful hierarchical feature extraction and learning capabilities to make breakthroughs in the performance of object detectors. Meanwhile, the large?scale datasets and the tremendous improvement in computing power have also contributed to the vigorous development in this field. In this paper, the existing research of object detectors based on deep learning are reviewed in detail. First, we review the traditional object detection algorithms and its problems. Then, object detectors based on deep learning are introduced, and the region?based and single?stage benchmark detectors are summarized. After that, the current mainstream object detectors are concluded from eight perspectives of feature maps, context information, bounding box optimization, regional proposal, category imbalance processing, training strategy, weakly supervised learning and unsupervised learning. Finally, the problems to be solved in the object detectors are proposed and future research directions are prospected.

Key words: computer vision, deep learning, object detection, convolutional neural network

中图分类号: