在线咨询
中国工业与应用数学学会会刊
主管：中华人民共和国教育部
主办：西安交通大学
ISSN 1005-3085  CN 61-1269/O1
工程数学学报

工程数学学报 ›› 2024, Vol. 41 ›› Issue (3): 410-420.doi: 10.3969/j.issn.1005-3085.2024.03.002

• • 上一篇    下一篇

带有特征信息卷积神经网络的人脸识别算法

岳  也,  温瑞萍,  王川龙   

  1. 太原师范学院智能优化计算与区块链技术山西省重点实验室，晋中 030619
  • 收稿日期:2021-04-25 接受日期:2023-03-20 出版日期:2024-06-15 发布日期:2024-08-15
  • 通讯作者: 温瑞萍 E-mail: wenrp@163.com
  • 基金资助:
    国家自然科学基金 (12371381)；山西省科技创新人才团队建设重点项目 (202204051002018).

Face Recognition Algorithm Based on Convolutional Neural Network with Feature Information

YUE Ye,  WEN Ruiping,  WANG Chuanlong   

  1. Shanxi Key Laboratory for Intelligent Optimization Computing and Block-chain Technology, Taiyuan Normal University, Jinzhong 030619
  • Received:2021-04-25 Accepted:2023-03-20 Online:2024-06-15 Published:2024-08-15
  • Contact: R. Wen. E-mail address: wenrp@163.com
  • Supported by:
    The National Natural Science Foundation of China (12371381); the Special Fund for Science and Technology Innovation Teams of Shanxi Province (202204051002018).

摘要：

图像分类中，卷积神经网络在人脸识别中取得了较大的进展。在卷积提取人脸图像特征信息操作时，当卷积核数目有限的情况下，可能提取到的特征值，如头发、纹理等，并不能很好的代表该人的主要特征，从而导致识别率降低，而增加卷积核数目又会导致识别时间增加。针对这一问题，提出了一种基于特征信息卷积神经网络的人脸识别方法。该方法在图像处理过程中，使用奇异值分解，选取前4个奇异值代表人脸的主要特征，快速滤除大部分无用的特征信息，形成新的图像特征模板库。利用卷积网络在提高网络感受野的同时不丢失特征图信息的优势，融合最具有代表性的特征信息，最大程度地捕捉图像信息。采用卷积神经网络模型和基于奇异值分解的特征融合的结构模型实现人脸识别，仿真实验结果表明，这种方法减少了算法的训练时间，提高了人脸识别的准确性。

关键词: 人脸识别, 奇异值分解, 特征提取, 卷积神经网络, 人脸数据库, 仿真实验

Abstract:

In image classification, convolution neural network has made great progress in face recognition. When convolution is used to extract face image feature information, when the number of convolution kernels is limited, the feature values, such as hair, texture, may not represent the main features of the person well, resulting in the reduction of recognition rate. To solve this problem, a face recognition method based on feature information convolution neural network is proposed in this paper. In the process of image processing, ingular value decomposition is used to select the first four singular values to represent the main features of the face, and most of the useless feature information is quickly filtered out. The convolution network can improve the receptive field of the network without losing the information of the feature map, and fuse the most representative feature information. The convolutional neural network model and the structural model of feature fusion based on singular value decomposition are used to realize face recognition. The simulation results show that this method reduces the training time of the algorithm and improves the accuracy of face recognition.

Key words: face recognition, singular value decomposition, eigenvalue extraction, convolu-tional neural network, face database, simulation experiment

中图分类号: