基于FPGA的人脸识别算法的设计与实现

发布时间：2017-12-27 15:35

本文关键词：基于FPGA的人脸识别算法的设计与实现　出处：《中国科学技术大学》2016年硕士论文　论文类型：学位论文

【摘要】：在信息爆炸式增长的当今,信息安全已经成为一个非常重要的课题,而人脸识别是生物特征识别的重要识别方法之一,相对于其他识别方法,比如指纹识别、基因识别、语音识别以及虹膜识别等,人脸识别具有不需要与人接触,操作简单、可跟踪以及识别高等优点,因而被广泛应用于国家安全、安保、出入口检测等领域。传统PC端的人脸识别方法已经比较成熟,但随着电子消费等概念的诞生,这些方法都需要移植到嵌入式的设备中。PC机一般不能满足嵌入式的要求,而DSP受到资源的限制,不能满足高速视频流的数据处理。而近二十年来,FGPA得到了高速的发展,性能和集成度不断提高,使得FPGA处理图像和视频成为可能。FPGA具有逻辑资源多,运算能力强,速度快,灵活性和可移植性等特点,而且目前很多家厂商都提供了FPGA+ARM的嵌入式设计,用户可以很容易的实现软硬件一体的协同设计。使用FPGA对图像和视频进行处理已成为一种趋势。比较通用的人脸识别的过程主要有人脸检测与定位、图像预处理、人脸特征提取以及人脸匹配等几部分。在人脸的检测与定位部分,本文主要根据肤色这一显著特征来实现人脸区域的分割,并通过形态学滤波等手段实现对图像中的人脸进行定位。图像预处理部分则实现了高斯滤波、中值滤波、图像的直方图均衡化算法以及canny边缘检测算子。在人脸特征提取部分,首先用MATLAB软件实现了主成分分析、独立成分分析、非负矩阵分解以及离散余弦变换四种特征提取方法,并将这些算法应用于ORL人脸数据库,其次分析了这些算法提取的特征向量维度以及训练集的大小对最终人脸识别率的影响,对比四种算法,离散余弦算法对ORL的数据库识别效果最好,识别率最高能达到97.5%,最后对DCT算法进行了FPGA的移植。将FPGA中提取的人脸特征向量通过AXI总线传到ARM中,并与ARM中的人脸数据库进行人脸匹配。本文最后在ZYNQ平台上实现了一个简要的人脸识别系统,该系统对实验室的人脸进行识别,系统最终能够达到91.6%的识别率。
[Abstract]:In the information explosion today, information security has become a very important topic, and face recognition is one of the important recognition method of biometric recognition, compared with other identification methods, such as fingerprint identification, gene recognition, speech recognition and iris recognition, face recognition has no need to contact with people, simple operation, can be tracking and recognition of the advantages, so it has been widely used in national security, security, access detection etc.. The traditional PC face recognition method is more mature, but with the birth of the concept of electronic consumption, these methods need to be transplanted into embedded devices. PC can not meet the requirements of embedded system, but DSP is limited by resources and can not meet the data processing of high speed video stream. In the past twenty years, FGPA has been developing rapidly, and its performance and integration are increasing, making it possible for FPGA to deal with images and video. FPGA has many advantages, such as many logical resources, strong computing power, fast speed, flexibility and portability, and many manufacturers have provided FPGA+ARM's embedded design. Users can easily achieve collaborative design of hardware and software. The use of FPGA to process images and video has become a trend. The general process of face recognition mainly includes face detection and location, image preprocessing, face feature extraction and face matching. In the part of face detection and location, this paper mainly realizes the segmentation of face regions according to the salient feature of skin color, and realizes the location of faces in images by morphological filtering. The image preprocessing part realizes the Gauss filter, median filter, image histogram equalization algorithm and Canny edge detection operator. In the part of extracting face features, firstly, using MATLAB software, principal component analysis, independent component analysis, non negative matrix factorization and discrete cosine transform four feature extraction methods, and these algorithms are applied to ORL face database, then analyzes these algorithms to extract the feature vector dimension and the size of the training set of end face the recognition rate, the comparison of four algorithms, the best database recognition algorithm for ORL discrete cosine, the highest recognition rate can reach 97.5%, and finally the DCT algorithm is the transplantation of FPGA. The face feature vectors extracted from FPGA are passed through the AXI bus to ARM, and the face database is matched with the face database in ARM. Finally, a brief face recognition system is implemented on the ZYNQ platform. The system can recognize the face of the laboratory, and the system can reach 91.6% recognition rate.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【参考文献】