摘要
在基于区域的卷积神经网络提出后,深度卷积网络开始在目标检测领域普及,更快的基于区域的卷积神经网络将整个目标检测过程合成在一个统一的深度网络框架上.随后YOLO和SSD等目标检测框架的提出进一步提升目标检测的效率.文中系统总结基于深度网络的目标检测方法,归为2类:基于候选窗口的目标检测框架和基于回归的目标检测框架.基于候选窗口的目标检测框架首先需要在输入的图像上产生很多的候选窗口,然后对这些候选窗口进行判别.这里的判别包括:对窗口包含物体的类别(包括背景)进行判断、对窗口的位置进行回归.基于回归的目标检测方法将图像目标检测看作是一个回归的过程.在此基础上,在PASCAL_VOC和COCO等主流数据库上对比目前两类目标检测框架中的主流方法,分析两类方法各自的优势.最后根据当前深度网络目标检测方法的发展趋势,对目标检测方法未来的研究热点做出合理预测.
Deep convolutional network is prevalent in object detection task. Region-based convolutional neural network（ RCNN） bridges the gap between the classification of deep convolutional network and the object detection task well. Then the whole object detection process is aggregated into a unified deep framework by Faster-RCNN. You only look once（ YOLO） and single shot multibox detector（ SSD） effectively improve the efficiency of object detection. Different deep object detection frameworks are comprehensively analyzed and divided into two categories： the proposal based framework and the regression based framework. The proposal based framework is utilized to generate thousands of candidate proposals and then classification and bounding box regression are conducted on these proposals. The regression based framework outputs the bounding box position through some special iterations directly.Furthermore, the advantage for different kinds of frameworks is demonstrated through adequate experiments on the mainstream database like PASCAL_ VOC and COCO. Finally,the development direction of object detection is discussed.
作者
吴帅
徐勇
赵东宁
WU Shuai1, XU Yong1 , ZHAO Dongning1, 2(1. IntelliSense and Bioinformatics Innovation Team, HIT Institute of Technology Shenzhen Graduate School, Harbin Institute of Technology, Shenzhen 518055; 2. College of Information Engineering, Shenzhen University, Shenzhen 51800)
出处
《模式识别与人工智能》
EI
CSCD
北大核心
2018年第4期335-346,共12页
Pattern Recognition and Artificial Intelligence
