基于深度学习的人脸识别技术研究

发布时间：2018-02-01 14:53

本文关键词： 深度神经网络人脸识别深度多模型融合卷积神经网络组合特征　出处：《西南交通大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着人工智能的快速发展,如何准确、有效的识别用户身份,提升信息安全成为一项重要的研究课题。相较于传统的卡片识别、指纹识别和虹膜识别,人脸识别具有许多优点。它的非接触性、非强制性和并发性,易被用户所接受,已广泛应用于教育、电子商务等多个领域。深度学习是机器学习领域中的新兴分支之一。与传统浅层网络不同,深度学习受人脑工作机制启发,构建了深层网络结构和相应的训练方法。深度卷积神经网络(deep convolution neural networks,DCNN)源于多层前向网络,经过不断发展,已成为当前图像识别领域的研究热点。它依靠深层非线性网络结构和大规模的训练数据,实现复杂函数逼近,从而获得更本质和鲁棒的图像特征,有效提升了后续分类与识别的效果。近年来随着深度卷积神经网络的引入,人脸识别的准确率得以跨越式提升。然而,不同模型的训练集和网络结构差异较大,使得每个模型都有各自的特点。对此,本文研究了一种基于深度多模型融合的人脸识别方法,通过融合多个人脸识别模型提取的特征构成组合特征,再利用深度神经网络训练组合特征构建人脸识别分类器,可以得到融合多个模型优点的改进模型。主要的工作如下:(1)分析和对比基于卷积神经网络且开源的人脸识别算法,通过实验筛选了 2种基础模型。对基础模型提取的基础特征进行降维、归一化、融合,得到组合特征,作为后续深度神经网络的输入。(2)构建基于深度多模型融合的深度神经网络,训练组合特征,获得融合不同基础模型优点的改进模型。(3)进一步分析改进模型并设计了多组实验,包括不同的训练集、DNN参数和基础特征权重。统计了基础模型和改进模型在LFW数据集上的详细测试数据,探索改进模型提升的原因。在采用较小数据集的情况下,本文方法在人脸识别权威测试集LFW和YTF上获得了 99.1%和93.32%的精度,相对于基础模型分别提高0.57%和0.52%。而且通过对LFW测试数据的进一步分析,探讨了改进模型在融合不同基础模型优点方面的有效性。
[Abstract]:With the rapid development of artificial intelligence, how to accurately and effectively identify users and improve information security has become an important research topic. Compared with traditional card recognition, fingerprint recognition and iris recognition. Face recognition has many advantages. Its non-contact, non-mandatory and concurrent, easy to be accepted by users, has been widely used in education. E-commerce and other fields. Deep learning is one of the new branches in the field of machine learning. Unlike the traditional shallow network, deep learning is inspired by the working mechanism of human brain. The deep convolution neural networks and deep convolution neural network were constructed. DCNN, which originates from multilayer forward network, has become a research hotspot in the field of image recognition through continuous development. It relies on deep nonlinear network structure and large-scale training data to achieve complex function approximation. In recent years, with the introduction of deep convolution neural network, the accuracy of face recognition can be improved by leaps and bounds. Different models have different training sets and network structure, which makes each model have their own characteristics. In this paper, a face recognition method based on depth multi-model fusion is proposed. The features extracted from multiple face recognition models are fused to form the combined features, and then the combined features are trained by the depth neural network to construct the face recognition classifier. The main work is as follows: 1) analyze and compare the face recognition algorithm based on convolution neural network and open source. Through experiments, two basic models are selected. The basic features extracted from the basic model are reduced, normalized, fused, and combined features are obtained. As the input of the subsequent depth neural network, we construct the depth neural network based on depth multi-model fusion, and train the combination features. An improved model combining the advantages of different basic models is obtained. (3) further analysis of the improved model and design of a number of experiments, including different training sets. DNN parameters and basic feature weights. The detailed test data of the basic model and the improved model on the LFW data set are analyzed to explore the reasons for the improvement of the improved model. In the case of using smaller data sets. In this paper, the accuracy of 99.1% and 93.32% is obtained on the face recognition authoritative test set LFW and YTF. Compared with the basic model, the improvement is 0.57% and 0.52 respectively. The effectiveness of the improved model in combining the advantages of different basic models is discussed through further analysis of the LFW test data.
【学位授予单位】：西南交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP181

【参考文献】