基于深度神经网络的人脸识别算法研究

发布时间：2017-12-27 15:33

本文关键词：基于深度神经网络的人脸识别算法研究　出处：《浙江大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着计算机互联网技术的进步及社会信息化程度的提升,如今的各个领域都对身份识别技术提出了越来越高的需求。人脸识别技术因为人脸稳定且方便检测的特性而得以被广泛应用于金融、安防、电子商务、手机等领域。因为拥有巨大的应用前景,人脸识别也成为了计算机视觉、模式识别领域中的热门课题之一。而深度学习和神经网络的发展带动了人脸识别技术的进步。不断有人提出新的深度学习模型,这些模型也被用于人脸识别,并达到了非常好的效果。本文将一个具有多层卷积的深度神经网络用于提取人脸特征,该模型同时利用了输入的识别信号和验证信号来训练模型,并通过PCA降维、联合贝叶斯等算法对提取得到的特征进行人脸验证。本文提出对卷积神经网络中的pooling层进行改进,将其替换为卷积操作或全连接操作,提高了模型获取信息的能力。同时本文提出了基于人脸关键点位置的模型,在一次训练过程中可以利用不同关键点周围的信息,将单个模型的准确率提高到99%以上,解决了多模型训练耗时的问题。使用不到70万张的人脸图像和输入为100的图像大小,我们通过对模型的优化在LFW数据集实现了单模型99.17%的平均准确率,仅仅使用单模型的效果就接近于多模型融合。
[Abstract]:With the progress of computer Internet technology and the improvement of social information level, nowadays, more and more demands for identity recognition technology have been put forward in various fields. Face recognition technology has been widely used in the fields of finance, security, electronic commerce, mobile phone and so on, because of the characteristics of stable and convenient detection. Face recognition has become one of the hot topics in the field of computer vision and pattern recognition because of its great application prospect. And the development of deep learning and neural network has led to the progress of face recognition technology. A new model of deep learning has been proposed, and these models are also used for face recognition and have achieved good results. In this paper, a multi-layer convolution deep neural network is used to extract facial features. This model uses the input recognition signal and verification signal to train the model, and uses PCA dimensionality reduction and joint Bias algorithm to verify the extracted features. This paper proposes an improvement on the pooling layer in the convolution neural network, which is replaced by a convolution operation or a full connection operation, which improves the ability of the model to obtain information. At the same time, a model based on the location of the key points is proposed in this paper. In a training process, the information around different key points can be used to improve the accuracy of a single model to more than 99%, which solves the problem of multi model training time consuming. Using less than 700 thousand face images and 100 image size, we achieved the average accuracy of single model 99.17% in LFW dataset by optimizing the model. The effect of using single model is close to multi model fusion.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【相似文献】