基于Hadoop的多姿态人脸识别

发布时间：2018-03-28 02:37

本文选题：Hadoop　切入点：多姿态人脸识别　出处：《吉林大学》2017年硕士论文

【摘要】：随着社会的不断发展,科技水平的不断进步,人脸识别技术由于具有结果直观、隐藏性好、操作方便的优越性,被广泛应用于信息安全、金融安全、反恐环境、刑侦调查等领域。但是在非约束环境下,摄像头往往采集不到非常合适的人脸,比如面部遮挡,复杂环境干扰,表情、姿态的变化,都会在一定程度上导致人脸识别效果下降。另外,伴随着不断扩大规模的数据生成点,图像数据年都以20%的增长率快速增加,如何在庞大的数据中快速的查找到目标人脸图像,这就需要将云计算服务应用在传统的识别技术中。所以,为了解决以上出现的问题,本文以Hadoop结构为基础,构建了云计算平台,并在Hadoop云计算平台下对非约束环境中多姿态人脸识别展开了研究。本文主要从以下几个方面展开研究工作:1、研究了目前流行的Hadoop云平台结构,以VMware workstation作为虚拟机,创建了3个ubuntu系统,并将3个ubuntu系统组成Master/Slaver结构,同时配置好相关的文件、网络和软件,最后构建Hadoop完全分布式系统。2、研究了传统的神经网络的组成与特点,然后进一步依据神经网络特点分析卷积神经网络的结构,并且主要对卷积神经网络的卷积层、池化层和激活层进行了进一步的研究,最后简要总结了卷积神经网络的特点。3、通过对当前流行的Le Net-5卷积神经网络模型进行性能试验与分析,研究卷积神经网络的卷积核尺寸与数目、正则化方式、激活函数、池化方式等参数对该模型相关性能的影响,从而能够选取最合适的模型参数来构建出最优的卷积神经网络模型,并且在CAS-PEAl人脸库上进行了验证。4、利用本文搭建的Hadoop云平台进行图片集的收集与整理,然后使用本文改进的卷积神经网络对收集到的图片集合进行特征提取,并通过PCA算法将提取的特征向量进行降维,最后识别阶段采用余弦相似度度量算法进行目标人脸识别。本实验在CAS-PEAl人脸库进行,并对本文改进的多姿态人脸识别算法在识别正确率与识别时间上进行了有效的分析。
[Abstract]:With the development of society and the progress of science and technology, face recognition technology has been widely used in information security, financial security, anti-terrorism environment because of its advantages of intuitive results, good concealment and convenient operation. But in unconstrained environments, cameras often fail to capture very suitable human faces, such as facial occlusion, complex environmental disturbances, facial expressions, changes in posture, In addition, with the increasing scale of data generation point, the annual growth rate of image data increases rapidly by 20%, how to find the target face image quickly in the huge data. Therefore, in order to solve the above problems, this paper constructs cloud computing platform based on Hadoop structure. The multi-pose face recognition in unconstrained environment is studied on the Hadoop cloud computing platform. This paper mainly studies the structure of the popular Hadoop cloud platform in the following aspects: 1. VMware workstation is used as the virtual machine. Three ubuntu systems are created, and three ubuntu systems are formed into Master/Slaver structure. At the same time, the related files, network and software are configured. Finally, a fully distributed Hadoop system .2is constructed, and the composition and characteristics of traditional neural networks are studied. Then the structure of convolution neural network is analyzed according to the characteristics of neural network, and the convolution layer, pool layer and activation layer of convolutional neural network are studied further. Finally, the characteristics of convolution neural network are summarized briefly. Through the performance test and analysis of the current Le Net-5 convolution neural network model, the size and number of convolution cores, regularization mode, activation function of the convolutional neural network are studied. The influence of the parameters such as pool mode on the performance of the model can be used to select the most suitable model parameters to construct the optimal convolution neural network model. And on the CAS-PEAl face database validation. 4, using the Hadoop cloud platform built in this paper to collect and organize the picture set, and then use the improved convolution neural network to extract the features of the collected image set. The extracted feature vector is reduced by PCA algorithm, and the cosine similarity measure algorithm is used in the final recognition stage. This experiment is carried out in the CAS-PEAl face database. And the improved multi-pose face recognition algorithm is analyzed effectively in the recognition accuracy and recognition time.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】