基于动态记忆网络的智能视觉问答系统的设计与实现  

Design and Implementation of Intelligent Visual Question Answering System Based on Dynamic Memory Network

作　　者：吴玥 高会议[2] 陈雷[2] 曾明昭 万莉 WU Yue;GAO Huiyi;CHEN Lei;ZENG Mingzhao;WAN Li(Institute of Physical Science and Information Technology,Anhui University,Hefei 230601,China;Institute of Intelligent Machines,Chinese Academy of Sciences,Hefei 230031,China;University of Science and Technology of China,Hefei 230026,China)

机构地区：[1]安徽大学物质科学与信息技术研究院,安徽合肥230601 [2]中国科学院合肥智能机械研究所,安徽合肥230031 [3]中国科学技术大学,安徽合肥230026

出　　处：《仪表技术》2020年第5期18-22,共5页Instrumentation Technology

基　　金：国家自然科学基金(31601268);国家自然科学基金(31771677)。

摘　　要：随着计算机技术和人工智能的发展,视觉问答在未来的导盲机器人等领域的应用会越来越广。为了探究这一应用的可行性和更高效地收集视觉问答的数据资源,研究构建了一种基于动态记忆网络的智能视觉问答系统,允许对不同摄像头中的图像或本地输入的图像进行提问,系统会自动地进行处理并给出较为合理的回答。另外,该系统也实现了将人为校验和修正后的视觉问答数据实时地储存到数据库中,为以后的工作提供数据支持。实验结果表明,该智能视觉问答系统在两个不同的测试集上的准确率可以达到60%左右,说明系统是可行有效的。With the development of computer technology and artificial intelligence technology,the visual question an-swering system will be more and more widely used in the fields of blind guide robots in the future.In order to explore the feasibility of this application and collect data resources for visual question answering more efficiently,this study develops an intelligent visual question answering system based on dynamic memory network.This system allows user to ask questions about images in different cameras or local input images,and then it will give a more reasonable answer automatically.Be-sides,the system can also store the correct visual question answering data in the database in real time,which could pro-vide data support for future work.The experimental results show that the accuracy of the intelligent visual question answer-ing system can reach about 60%on two different test sets,indicating that the system is feasible and effective.

关 键 词：视觉问答 动态记忆网络 智能系统 数据库 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]