摘要
当前大多数深度强化学习方法在目标检测方面的召回率较低。为此,提出一种层级偏移的动态搜索方法。在原有层级搜索的基础上,采用了锚点的思想,增加区域偏移,避免层级产生的区域局限,使得搜索更加灵活。结合Double DQN与Dueling DQN的优势,以Double Dueling DQN的网络结构作为智能体深度增强网络的结构。实验结果表明,与原有层级搜索方式相比,该方式的目标检测的精确度与召回率较高。
In order to solve the problem of low recall rate in object detection with the deep reinforcement learning method,on the basis of simulating human visual mechanism,a dynamic searching hierarchical offset method is proposed.It uses the idea of anchors based on the original hierarchical searching method,which adds a region offset.This method avoids the limitations generated by hierarchical searching method,and makes the search more flexible.This paper combines the advantages of Double DQN and Dueling DQN,using Double Dueling DQN network structure as the deep reinforcement learning network of the agent.Experimental results show that the accuracy and recall ratio are higher than the original hierarchical searching method.
作者
秦升
张晓林
陈利利
李嘉茂
QIN Sheng 1,2,ZHANG Xiaolin 1,CHEN Lili 1,LI Jiamao 1(1.Shanghai Institute of Micro System and Information Technology,Chinese Academy of Sciences,Shanghai 200050,China; 2.University of Chinese Academy of Sciences,Beijing 100049,Chin)
出处
《计算机工程》
CAS
CSCD
北大核心
2018年第6期253-258,共6页
Computer Engineering
基金
上海市科学技术委员会科研计划项目(17YF1427300)
关键词
人类视觉机制
深度强化学习
层级偏移
目标检测
马尔科夫决策过程
human visual mechanism
deep reinforcement learning
hierarchical offset
object detection
Markov Decision Making Process（MDP）
