«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

30

期数:

2020年07期

页码:

1-5

栏目:

智能、算法、系统工程

出版日期:

2020-07-10

文章信息/Info

Title:

An FPGA-based SOC Design of Dynamic Loading of Convolution Neural Network

文章编号:

1673-629X(2020)07-0001-05

作者:

许永全; 冯玉田

上海大学 通信与信息工程学院,上海 200444

Author(s):

XU Yong-quan; FENG Yu-tian

School of Communication and Information Engineering,Shanghai University,Shanghai 200444,China

关键词:

机器视觉; FPGA; 卷积神经网络; SOC 系统; MicroBlaze 处理器

Keywords:

machine vision; FPGA; convolutional neural network; SOC system; MicroBlaze processor

分类号:

TP39

DOI:

10. 3969 / j. issn. 1673-629X. 2020. 07. 001

摘要:

机器视觉是人工智能与现代工业应用结合的结果,通过机器视觉将目标物品转换成具有特定含义的数字信号,进而根据信号的信息测试目标物品并控制现场设备的动作,实现使用机器代替人的目标。 文中设计了一种基于 FPGA 的SOC 软硬件结合的系统,解决了以卷积神经网络为基础的机器学习算法在实际工业应用的具体问题。 该系统在 FPGA 内部例化了双核 MicroBlaze 处理器,集成了 DLA(deep learning accelerator) 卷积神经网络硬件算法核心,能够实时高速采集图像和实时分析处理;同时还设计了一套软件系统,帮助算法工程师实现自己的网络模型。 针对复杂多变的现场应用,设计了一套支持网络动态加载的系统环境,提高了在机器视觉领域的适用性,尤其适用于以卷积神经网络为基础的工业自动化领域,如机器视觉检测、结构光扫描测量、机器人引导等。 实验结果表明,系统在充分发挥 FPGA 计算速度优势的同时, 简化了应用工程师的设计过程,有效降低了实际应用的实施难度。

Abstract:

Machine vision is the result of artificial intelligence and modern industrial applications. The target objects are converted into digital signals with special meaning by machine vision,and then according to the information of the signals,the target objects are tested and the device actions in the control field are controlled to achieve the goal of replacing human beings with machines. We design a FPGA-based SOC software and hardware combination system to solve the specific problems of machine learning algorithm based on convolution neural network in practical industrial application. The system instantiates a dual-core MicroBlaze processor in FPGA and integrates DLA(deep learning accelerator) hardware algorithm core of convolution neural network,which is capable of real-time high-speed image acquisition and real-time analysis and processing. At the same time,a software system is designed to help algorithmic engineer to realize his own network model. A system environment supporting network dynamic loading at complex and changeable field applications is designed to improve the applicability in the field of machine vision,especially suitable for industrial automation based on convolution neural network,such as machine vision detection, structured light scanning measurement, robot guidance and so on. The experiment shows that this system not only gives full play to the advantages of the computing speed of the FPGA,but also simplifies the design procedure for application engineers and effectively reduces the difficulty of practical applications.

常用功能

更新日期/Last Update: 2020-07-10