基于深度信念网络的人脸识别

发布时间：2018-06-26 04:35

本文选题：人脸识别 + 深度学习　；参考：《大连海事大学》2017年硕士论文

【摘要】：随着我们周边生活信息化,自动化的不断加深,监控摄像头已经变得随处可见,有些地方还出现了基于指纹和人脸识别的门禁系统。在这些系统中人脸识别技术都有重要的应用,人脸识别是生物特征识别的一个重要部分,是一门涵盖了计算机,图像处理,信号处理,模式识别,心理学等多种学科的技术。虽然人脸识别的研究以有很长的历史但仍有极高的研究价值。深度信念网络(Deep Belief Network,简称DBN)和对比散度算法由学者Geoffrey Hinton在2006年提出。这开启了深度学习的新时代,在现在深度学习与大数据相结合给人们设计出真正的人工智能指出了一条可能的道路。人脸是一个复杂系统,浅层学习在面对复杂系统时会出现数据过拟合等问题,而采用多层网络结构可以有效提升识别效果,本文采用深度信念网络。本文研究的主要内容如下。1,利用深度信念网络和对比散度算法设计和训练人脸识别系统,实验表明基于深度信念网络的人脸识别系统能够有效地完成分类任务。2,在设计网络时就深度信念网络层数,训练网络时参数等会对识别效果产生影响的因素加以分析和讨论,得出预训练相关参数和网络层数的较优值。3在CMUPIE,Yale,UMIST人脸数据库上验证本文识别系统有效性,并就光照,分辨率,表情,姿态变化对识别效果影响情况进行实验测试,并与传统的人脸识别方法进行对比,测试结果表明本文的识别算法对人脸的识别率能达到90%以上,与传统算法相比识别率可以提高1%到2%。
[Abstract]:With the informatization of our peripheral life and the deepening of automation, surveillance cameras have become ubiquitous, and in some places, access control systems based on fingerprint and face recognition have emerged. Face recognition is an important part of biometric recognition, which includes computer, image processing, signal processing, pattern recognition, psychology and so on. Although the research of face recognition has a long history, it is still of great value. Deep belief Network (DBN) and contrast divergence algorithm were proposed by Geoffrey Hinton in 2006. This opens a new era of deep learning, and now the combination of deep learning and big data points out a possible way to design true artificial intelligence. Face is a complex system. In the face of complex system, shallow learning will have problems such as data over-fitting, and the multi-layer network structure can effectively improve the recognition effect. In this paper, the depth belief network is used. The main contents of this paper are as follows: 1. Using depth belief network and contrast divergence algorithm to design and train face recognition system. The experiments show that the face recognition system based on depth belief network can effectively complete the classification task. The factors that affect the recognition effect are analyzed and discussed in the design of the network, such as the number of depth belief network layers and the parameters when training the network. The optimal value of pre-training parameters and the number of network layers is obtained to verify the effectiveness of this recognition system on CMUPIEE / UMIST face database, and the effect of illumination, resolution, facial expression and attitude changes on the recognition effect is tested experimentally. Compared with the traditional face recognition method, the test results show that the recognition rate of this algorithm can reach more than 90%, compared with the traditional algorithm, the recognition rate can be increased by 1% to 2%.
【学位授予单位】：大连海事大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】