计算机工程与应用 ›› 2020, Vol. 56 ›› Issue (15): 215-220.DOI: 10.3778/j.issn.1002-8331.1907-0011

• 图形图像处理 • 上一篇    下一篇

融合LeNet-5和Siamese神经网络模型的人脸认证算法研究

厍向阳,刘巧,叶鸥   

  1. 西安科技大学 计算机科学与技术学院,西安 710054
  • 出版日期:2020-08-01 发布日期:2020-07-30

Research on Face Verification Algorithm Based on LeNet-5 and Siamese Neural Network Model

SHE Xiangyang, LIU Qiao, YE Ou   

  1. College of Computer Science and Technology, Xi’an University of Science and Technology, Xi’an 710054, China
  • Online:2020-08-01 Published:2020-07-30

摘要:

基于人脸信息的身份认证对于个人安全和社会稳定都具有非常重要的意义。传统的人脸认证方法依赖人工构造视觉特征,易受外界条件影响,识别精度不高。深度学习模型以自主学习方式进行特征提取,能从复杂的数据中提取到人脸的隐性特征。然而大部分深度学习人脸认证方法需大量带有身份标记的训练样本,额外增加了标记数据的成本。针对以上问题,提出了融合LeNet-5和Siamese神经网络模型的人脸认证算法。该算法在Siamese神经网络框架基础上,引入LeNet-5卷积神经网络,将单分支LeNet-5卷积网络扩充为结构相同且参数共享的双分支LeNet-5卷积网络,通过缩小卷积核、增加卷积层来调整网络结构,使用Contrastive Loss函数对融合网络进行训练。实验结果表明,该算法在不同的人脸数据集上,均获取较高的识别精度。

关键词: 人脸验证, 深度学习, Siamese网络

Abstract:

Face-based identity verification is very important for personal security and social stability. The traditional face verification method relies on manual construction of visual features and is susceptible to external conditions, so the accuracy of face recognition is not high in real scenes. The deep learning model extracts features in autonomous learning and can extract the hidden features of the face in complex data. However, the most deep learning face verification methods require a large number of training samples with identity tags, which additionally increases the cost of tagging data. To solve these problems, this paper proposes a face verification method based on LeNet-5 and Siamese neural network. Based on the framework of Siamese neural network, this algorithm introduces LeNet-5 convolution neural network. It extends the single-branch LeNet-5 network to a double-branch LeNet-5 network with the same structure and shared parameters. It adjusts the network structure by reducing the convolution kernel and increasing the number of convolution layers. It uses Contrastive Loss function to train fusion network. Experimental results show that this algorithm can achieve high recognition accuracy on different face data sets.

Key words: face verification, deep learning, Siamese network