«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

期数:

2025年06期

页码:

18-26

栏目:

媒体计算

出版日期:

2025-06-10

文章信息/Info

Title:

Face Recognition Algorithm Based on MAML Online Meta-learning

文章编号:

1673-629X(2025)06-0018-09

作者:

范振雄; 邓春华

武汉科技大学 计算机科学与技术学院,湖北 武汉 430065

Author(s):

FAN Zhen-xiong; DENG Chun-hua

School of Computer Science & Technology,Wuhan University of Science and Technology,Wuhan 430065,China

关键词:

人脸识别; 元学习; 机器学习; 非限制性场景; 先验知识

Keywords:

face recognition; meta-learning; machine learning; non-restrictive scenarios; prior knowledge

分类号:

TP391.4

DOI:

10.20165/j.cnki.ISSN1673-629X.2025.0016

摘要:

深度学习在人脸识别中虽表现出色,但其性能受限于单个身份的训练样本的数量。 在非限制性场景下,人脸姿态多变,单个身份的注册样本不足时会导致识别效果不佳。 实际应用中,人脸数据常常稀少,有时每个身份仅有一张照片,传统机器学习方法难以处理这种情况。 因此,在样本稀缺时提高人脸识别的准确性和适应性是很大的挑战。 为解决上述问题,该文提出一种基于元学习的少量样本学习算法。 此算法运用从大规模人脸数据集中获得的先验知识,在注册样本稀少(甚至仅一张)时能进行有效推理,解决了少量样本甚至单样本人脸图像在传统机器学习方法下分类效果不佳的难题。 实验表明,该方法能提升不同模型的识别准确率,并在各种真实数据集上展现高效识别能力。 在正脸、侧脸、光照变化及配饰遮挡等多种真实场景的数据集上,该方法均展现出了卓越的识别能力。

Abstract:

Although deep learning performs well in face recognition,its performance is limited by the number of training samples for a single identity. In the non - restrictive scenario, the face posture is changeable, and the registration sample of a single identity is insufficient,which will lead to poor recognition effect. In practice,face data is often scarce,and sometimes there is only one photo for each identity. Traditional machine learning methods struggle to handle this situation. Therefore,it is a great challenge to improve the accuracy and adaptability of face recognition when samples are scarce. In order to solve the above problems,we propose a learning algorithm based on meta-learning with a small number of examples. The algorithm uses prior knowledge gained from large-scale face datasets to make effective inferences when there are few registered samples ( or even just one). This solves the problem that a small number of samples or even a single sample of face images do not perform well in classification under traditional machine learning methods. Experiments show that the proposed method can improve the recognition accuracy of different models and show efficient recognition ability on various real datasets. The proposed method shows excellent recognition ability on a variety of real-world scene datasets such as front face,side face,lighting change,and accessory occlusion.

常用功能

更新日期/Last Update: 2025-06-10