0DM642的嵌入式人脸识别系统的设计与实现

DM642的嵌入式人脸识别系统的设计与实现

摘 要

二十世纪末，随着计算机处理速度的飞速提高和图像处理算法的革命性改进，作为生物特征识别技术的一个重要研究方向——人脸识别技术在各种生物特征识别技术中脱颖而出，以其独特的主动性、非侵犯性和用户友好等众多优势受到世人瞩目。人脸识别技术无论在公安、军队、海关、交通、金融、社保、医疗及其他民用安全控制行业和部门都存在着广阔的发展空间。然而，人脸识别技术的产业化道路并不平坦，目前，其应用产品主要仍以基于PC平台的产品为主，很难满足各种情形下的应用需求。随着数字信号处理器（Digital Signal Processor, DSP）的处理速度和各项性能指标的大幅提升，用DSP来实现嵌入式人脸识别系统已经成为可能。

本课题源于上海银晨智能识别科技有限公司承担的上海市科委2010年世界博览会安防产品专项研究项目。全文从硬件平台、算法实现、程序优化等多方面详细论述了在TMS320DM642 DSP上实现人脸识别算法的过程，并将实现的嵌入式人脸识别系统与银晨公司上一代基于TMS320C6204 DSP的产品进行了比较。比较结果显示,我们设计的基于DM642的新系统在功率仅为原系统一半的情况下，识别速度比基于C6204的产品有10倍以上的提升。

在算法实现方面，我们根据基于DSP的嵌入式系统需要快速、高效算法的特点和要求选择了一种基于灰度图像横纹重心模板匹配的人脸检测算法。通过很多次的实验，我们对原有算法中的模板进行了改进，并且设计了灰度和边缘两种特征检查方案使整个人脸识别系统能够完成复杂背景下多级、多角度的人脸检测和识别。

DM642的嵌入式人脸识别系统的设计与实现

在硬件设计方面，考虑到以往的DSP芯片接外设能力较差、需要另配（如基于ARM处理器的）控制板来管理外部设备，为了提高系统的集成度，我们选择了美国德州仪器（TI）公司C6000系列中性能最强的定点视频专用DSP芯片——TMS320DM642，设计了一种DSP单处理器的硬件架构。由于DM642具有多个专用视频接口和网口等外设接口，使我们在用它实现人脸识别算法的同时能够对各种外部设备进行控制，大大降低了系统的体积和功耗。

我们设计的系统在实时性、识别率和功耗这三项指标的测评上都达到了嵌入式人脸识别系统的设计要求，完全可以满足产业化应用要求。

我们相信在相关技术，尤其是新的算法和芯片技术的推动下，在不久的将来，以人脸识别为代表的生物特征识别技术将从根本上改变人们的生活方式。

关 键 词：人脸识别，人脸检测，数字信号处理器，嵌入式系统，生物特征识别

DM642的嵌入式人脸识别系统的设计与实现

Abstract

Since the last decade of the 20th century, the prodigious development ofcomputer speed and the revolutionary improvement of digital imageprocessing had enabled the face recognition technology to distinguishitself from other biometrical identification technologies for its active,non-aggressive and user-friendly characteristics. Face recognitiontechnology is used in various applications and attracts more and moreattention around the world. Its possible application includes: socialsecurity, military, custom, traffic, finance and any public security relatedfield. Therefore, the way to industrialization remains difficult. Most of thecommercial products based on this technology are running on personalcomputers or mainframes. We are in lack of embedded systems that aresuitable for various applications. As the semiconductor industry boomedin the last few decades, the performance of DSP has been augmenteddramatically. To realize an embedded human face recognition system on aDSP has no longer been a dream.

This dissertation is a written testimony of the joint-project betweenSchool of Microelectronics of Shanghai Jiao Tong University andShanghai ISVISION Technology Co., Ltd. This paper covers the overalldesign procedure of an embedded face recognition system based onTMS320DM642 DSP including the hardware architecture, theimplementation of algorithm and the optimization of code. We alsopresent a comparison of our system with a TMS320C6204 basedembedded face recognition product of ISVISION. It is proved that oursystem is ten times superior to the old product in terms of recognitionspeed.

DM642的嵌入式人脸识别系统的设计与实现

As for the implementation of algorithm, we have chose a fast, efficienthuman face detection algorithm that is the “Gravity-center template basedhuman face feature detection” and added several templates to thealgorithm based on our experiments. We have also designed a grayscaleand edge feature detection method so as to adapt our system to detect faceon a more complex, multi-angle background.

We have designed a hardware platform which contains oneTMS320DM642 Video/Image DSP of the C6000 family of TEXASINSTRUMENT. We have no need of an ARM chip to manage theperipherals since the DM642 is an extra competent DSP at managing theperipherals such as video port and Ethernet interface. This reducestremendously the power consumption of the whole system.

Our system meets all three criteria of the design requirement of anembedded face recognition system. It is viable for the demand ofindustrialization.

We fully believe that the face recognition as the flag fleet of thebiometrical identification will change the way we live in the nearestfuture.

Keywords：Face Recognition, Face Detection, DSP, Embedded System,Biometric Identification

DM642的嵌入式 …… 此处隐藏：11137字，全部文档内容请下载后查看。喜欢就下载吧 ……