基于卷积神经网络的人脸验证研究

发布时间：2018-04-14 18:05

本文选题：人脸识别 + 人脸验证　；参考：《湘潭大学》2017年硕士论文

【摘要】：人脸验证是人脸识别研究问题中一个重点也是一个难点,因为最近几年因特网技术的发展,如何快速的进行身份验证以确保个人信息安全,已经成为一个热门话题。由于人脸验证问题是身份验证中比较重要的生物验证方法,从而使人脸验证成为了一个新的研究热点。人脸验证是一个二类验证问题即给定一张人脸图片把它和已知身份的人脸图片进行对比并判断两张图片是否是同一个人。本文主要是基于卷积神经网络研究的延展。在应对于人脸验证问题时分成了限制性条件和非限制条件下两种情形并且提出了两种基于卷积神经网络的人脸验证模型。主要的研究工作如下:1在基于Yale B数据库和AR数据库两个人脸数据库上构造了一个混合模型的卷积神经网络的人脸验证方法。相比于以前人脸验证方法,这种方法对人脸验证操作进行了分段操作。并且使用了PCA降维和SVM的验证分类操作。该方法相比于传统方法,在限制性实验环境下人脸验证的准确率有较好的提升。2在应对非限制条件下人脸图片时由于混合卷积神经网络的局限性,所以对上述的混合卷积神经网络进行了优化和改进从而构造了一个并行三通道卷积神经网络结构的人脸验证模型。3在LFW数据库上构造了两种不同的连接方式的三通道并行卷积神经网络模型。第一种采用了传统的全连接的方式。第二种是使用了局部连接的方式。使用两种连接的网络是为了比较在不同的连接方式下那种连接方式可以提高准确率。为了提高整个模型的正确率,模型进行了两次训练。第一次使用了SGD的优化函数进行首次训练,并保存下最好结果的模型。第二次使用了Adadelta优化函数在第一次训练好的模型上进行了二次训练以此来提高整个模型的准确率。
[Abstract]:Face verification is an important and difficult problem in face recognition, because with the development of Internet technology in recent years, how to quickly authenticate to ensure the security of personal information has become a hot topic.Face verification is an important biometric method in authentication, which makes human face verification become a new research hotspot.Face verification is a kind of verification problem that is to compare a face image with a face image of known identity and judge whether two images are the same person.This paper is mainly based on the extension of convolution neural network research.In this paper, the face verification problem is divided into two cases: restricted and unconstrained, and two kinds of face verification models based on convolution neural network are proposed.The main work of this paper is as follows: 1. Based on Yale B database and AR database, we construct a hybrid model based on convolution neural network for face verification.Compared with the previous face verification method, this method performs segmentation operation.And the verification classification operation of PCA reduction and SVM is used.Compared with the traditional method, the accuracy of face verification in the restricted experimental environment is better than that of the traditional one. 2. When dealing with face images under unconstrained conditions, the proposed method is limited by hybrid convolution neural networks.So the hybrid convolution neural network mentioned above is optimized and improved to construct a parallel three-channel convolutional neural network model of face verification. 3. 3 constructs two kinds of three different connection modes on LFW database.Channel parallel convolution neural network model.The first uses the traditional full-connected approach.The second is the use of local connections.Two kinds of connection networks are used to compare which connection modes can improve the accuracy of different connection modes.In order to improve the accuracy of the whole model, the model was trained twice.For the first time, the optimization function of SGD is used for the first time, and the best result model is saved.In the second time, the Adadelta optimization function is used to improve the accuracy of the model.
【学位授予单位】：湘潭大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【参考文献】