设计和实现基于DM642 DSP的人脸识别系统

上海交通大学硕士学位论文

设计和实现基于DM642 DSP的人脸识别系统

姓名：郑奕刚申请学位级别：硕士专业：信号与信息处理指导教师：陈健20050101

基于DM642 DSP的人脸识别系统的设计与实现

摘要

人脸检测和人脸识别是当前非常热门的研究领域

如何把人脸检测识别技术从实

验室转移出来

主流产品都是基于通用台式电脑的

当前这方面的

巨大的功耗

需要人力来管理等诸多不利条件都限制了这一技术的进一步产业化

作

者设计了一套采用高性能DSP的人脸识别系统有运算速度快

功耗低

DSP芯片具而德州仪器

把最

公司新推出的DM642芯片更是当前具有最高性能的DSP芯片高性能的芯片和当前的研究热点结合起来具有非常大的吸引力

如果有则进一

步确定人脸的大小和位置

在算法原理上它可以分为基于显式特征的方法和基于隐式特

征的方法

从算法上它可以

分为基于几何特征的识别方法和基于代数特征的识别方法两大类

它是一种基于积分图

具有非常稳定和高效的检测能力

在介绍了Adaboost人脸检测算法之后过程误合并的缺陷

阐述算法的实现

针对原算法对近距离人脸检测容易把两个脸错

人脸识别过程可以分为人脸检测和识别判断两个阶段

本文从算法背景然后

并且对识别的结果作了简要的分析

它离不开硬件平台的支持

本文从硬件和软件方面详细介绍了这套人脸识别

系统的设计思路和方法

在检测CIF大小的图像时

在做识别时

识别速度可达33帧每秒行的

紧贴应用的特点

由于它是与公司合作进

并得到了客户的赞赏和肯定

这次研究课题所产生的成果还引起了德州仪器公司的关注

TI Developer

Conference 2005

关键字

人脸识别

特征脸

　ＤＭ６４２

DESIGN AND IMPLEMENTATION OF

DM642 DSP BASED FACE RECOGNITION SYSTEM

ABSTRACT

Face detection and face recognition are the most challenging research areas in biometrics fields. It has been widely studied in recent years. While new algorithms and technologies are emerging out from labs, how to converter these interesting and exciting results into products become attractive. But the current applications are mostly based on general PC. They have some common disadvantages such as huge sizes, high power consumptions, unstable systems, and above all, high costs and prices, which prevented them from being widely used.

On the other hand, if these products are DSP-based, they will have exciting attributes. They can be little sized, power effective, stable and be embedded. So to combine the face detection and face recognition algorithms with high performance DSP chip is attractive. And the DM642 DSP chip is the most powerful DSP chip in the world, so we decide to design and implementation a DM642 DSP based face recognition system.

Face detection is to detect whether the input image contain any faces, if it

does, do further to find the positions and sizes of these faces. And face recognition is to decide who these faces belongs to respected. In this paper, the main attributes of face detection algorithms and face recognition algorithms was described and the currently development status was also touched.

Because this paper only used the famous adaboost face detection algorithm, the background of this algorithm was firstly given to show how the algorithm will do well in face detection. Then, the theory of the adaboost algorithm was deduced, it consists of two parts, training and detection. After that, how to implement the algorithm was represented by flow charts. Among this paper, there are many figures from which we can see the detection results of the algorithm. But it also has some unsatisfied attributes, when it was used to detect some coherent faces, it usually give error results. The author analyzed its cause and got rid of this problem by presenting a new merge algorithm.

Face recognition consists of face detection and recognition. To take the role of the 2nd step of face recognition, another well known algorithm, the eigenface based face recognition algorithm is approached in this paper. The eigenface based algorithm is similar to the PCA (principal component algorithm) algorithm. It is a powerful technique for extracting a structure from potentially high-dimensional data sets, which corresponding to

extracting the q eigenvectors which are associated with the largest q eigenvalues from the input distribution. If we firstly represent a face with a vector which consists of 576 data elements, but after adopting the eigenface based algorithm, we can represent the face with a far short vectors which only consists of very small number of data elements. For example, we can use only 3, 4 or 5 data elements to marker the face, which is the position of the face in the eigenface sub space. And we can use only these small sets of data to do further step during recognition process. In this paper, we showed many results of this algorithm.

As a DM642 DSP based face recognition, addition to the face detection and face recognition algorithms, how to design and implement a stable and high performance system was also at the center. The system consists of hardware and software. The hardware platform is the base of the system. The software consists of drive of peripherals and application software. By focusing on the hardware architecture and drive software, we described how to design the system.

DSP application system has its own attributes. In order …… 此处隐藏：6417字，全部文档内容请下载后查看。喜欢就下载吧 ……