基于嵌入式系统的人脸识别算法研究及其优化

发布时间：2018-05-18 10:18

本文选题：嵌入式系统 + 人脸识别　；参考：《杭州电子科技大学》2017年硕士论文

【摘要】：人脸识别作为一种友好的生物特征识别方式,具有不易伪造,容易获取、准确率高等优点。传统的人脸识别通常是在PC平台上实现的,近几年随着硬件性能的提升,嵌入式开发板逐渐用于实现人脸识别。由于嵌入式开发板的便携性好、稳定性高等优点,使嵌入式人脸识别系统的应用领域十分广泛。嵌入式开发板资源有限,可使用的算法具有一定的局限性。近几年随着人工智能技术的推广,高识别率的深度学习方法应用越来越广泛,但却无法直接运用在嵌入式开发板上。本文针对上述问题主要做了以下工作:(1)深入研究了嵌入式人脸识别的国内外发展和研究现状,总结了近几年深度学习的算法,并详细分析了其网络结构,结合嵌入式开发板的硬件资源有限的特点讨论了算法计算量。(2)研究了嵌入式人脸识别的各个组成部分。详细分析了基于嵌入式人脸识别常用的方法,讨论了常用的人脸数据库,研究了照片存储的格式,考虑了嵌入式人脸识别的耗时。(3)搭建了嵌入式人脸识别的开发环境。硬件方面,选择ARM架构的嵌入式开发板和摄像头;软件方面,选择开源的Linux操作系统。在PC机上进行了虚拟机的安装,建立了交叉编译环境,在开发板上进行了内核、驱动程序等相关的移植,为人脸识别应用程序的设计和开发搭建了一个稳定的运行环境。(4)提出了一种基于深度学习的pooling操作和DeepID算法的嵌入式系统pooling_patch算法。通过在ORL人脸库和本实验室成员的人脸图像库上进行实验,结果表明该方法不仅识别率高,且耗时短。
[Abstract]:As a kind of friendly biometric recognition method, face recognition has the advantages of difficult to forge, easy to obtain, high accuracy and so on. Traditional face recognition is usually implemented on PC platform. With the improvement of hardware performance, embedded development board is gradually used to realize face recognition in recent years. Because of the good portability and high stability of the embedded development board, the embedded face recognition system has a wide range of applications. The resources of embedded development board are limited, and the algorithm that can be used has some limitations. In recent years, with the popularization of artificial intelligence technology, the deep learning method with high recognition rate is more and more widely used, but it can not be directly used on the embedded development board. In this paper, the following work is done to solve the above problems: (1) the development and research status of embedded face recognition at home and abroad are deeply studied, the algorithms of depth learning in recent years are summarized, and the network structure of embedded face recognition is analyzed in detail. Combined with the limited hardware resources of the embedded development board, the computational complexity of the algorithm is discussed. (2) the components of embedded face recognition are studied. The common methods based on embedded face recognition are analyzed in detail, the commonly used face database is discussed, the format of photo storage is studied, and the time consuming of embedded face recognition is considered. Hardware, choose the ARM architecture of embedded development board and camera; software, choose the open source Linux operating system. The installation of virtual machine on PC, the establishment of cross-compiling environment, the transplantation of kernel, driver and other related programs on the development board are carried out. For the design and development of face recognition application, a stable running environment is built. (4) an embedded system pooling_patch algorithm based on deep learning pooling operation and DeepID algorithm is proposed. The experiments on ORL face database and our lab face image database show that this method not only has a high recognition rate but also takes a short time.
【学位授予单位】：杭州电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】