从RCNN到YOLO系列的物体检测系统综述  

Overview of Object Detection Systems from RCNNs to YOLOs

作　　者：王婉婷 姜国龙 褚云飞 陈业红[1] WANG Wan-ting;JIANG Guo-long;CHU Yun-fei;CHEN Ye-hong(School of Light Industry Science and Engineering,Qilu University of Technology(Shandong Academy of Sciences),Jinan 250353,China)

机构地区：[1]齐鲁工业大学(山东省科学院)轻工科学与工程学院,济南250353

出　　处：《齐鲁工业大学学报》2021年第5期9-16,共8页Journal of Qilu University of Technology

基　　金：国家自然科学基金青年基金(31901268);山东省自然科学基金(ZR2017LF028)。

摘　　要：作为计算机视觉的一个研究热点,物体检测近年来吸引了许多研究者的关注,在各领域的应用也呈现出蓬勃发展的势头。本文先对基于深度学习的两类物体检测框架,即基于候选区域(region proposal)的二步检测系统和非候选区域的一步检测系统进行综述。对于二步检测系统,本文介绍了RCNN(Region-based CNN)家族的算法及其发展;对于一步检测系统,主要介绍了YOLO(You Only Look Once)系列算法,包括各种算法的原理、创新点以及优缺点;接着再对数据集和检测算法的评价指标进行简单的阐述。最后,根据现有的物体检测算法所存在的局限及挑战对未来物体检测方法的方向进行展望。As a research hotspot of computer vision,object detection has attracted the attention of many researchers in recent years,which shows a booming momentum in various applications.This paper first reviews two types of object detection frameworks based on deep learning,i.e.,two-stage detection systems based on region proposal and one-stage systems with non-region proposal.For two-stage detection systems,the algorithms of RCNN(Region-based CNN)family and their development are given in this paper.For another,the YOLO series algorithm is investigated including the principle,innovation,merits and demerits in this paper.And then,the paper briefly describes the datasets and the evaluation metrics of detection algorithms.Finally,the future direction of object detection approaches is prospected according to its existing limitations and challenges.

关 键 词：物体检测 深度学习 二步检测系统 一步检测系统 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]