首页|采用特征增强和深度关系感知策略的3D人脸识别方法

采用特征增强和深度关系感知策略的3D人脸识别方法

扫码查看
针对2D人脸识别方法易受到外部环境干扰的问题,提出了一种基于深度学习的 3D人脸识别方法.该方法从人脸几何信息中获取特征,对光照等环境因素具有较强的鲁棒性.根据对现有研究内容的分析,设计了一个双域特征增强模块.该模块分别从通道域和空间域提取出人脸的局部特征,并将其作为全局特征的增强部分,从而获得更加完备的人脸特征.针对 3D人脸数据特性,提出了一种新的适合于3D人脸识别的特征学习策略.该策略旨在使人脸识别模型学习从 3D人脸的深度关系中提取身份特征,能够极大缓解三维人脸中噪声对特征计算的负面影响.通过实验,在公开数据集Bosphorus和Texas上分别获得了96.32%与98.93%的验证准确率,表明该方法能够获得更高的识别精度,并且在复杂情况下的人脸识别也具有一定优势.
A 3D Face Recognition Method Using Feature Enhancement and Depth Relationship Perception Strategy
To address the issue that 2D face recognition methods are susceptible to external environmental interference,a deep learning-based 3D face recognition method is proposed.The method extracts features from face geometric information and demonstrates strong robustness to environmental factors such as lighting.According to the analysis of existing research,a dual-domain feature enhancement module is designed.This module extracts local facial features from both the channel domain and the spatial domain,and uses them as enhancements to the global features,resulting in more comprehensive facial features.Additionally,a novel feature learning strategy tailored for 3D face recognition is proposed to address the characteristics of 3D face data.This strategy aims to enable face recognition models to extract identity features from the depth relationships of 3D faces and it can significantly alleviate the negative impact of noise in 3D faces on feature computation.On the public datasets Bosphorus and Texas,verification accuracies of 96.32%and 98.93%are achieved,respectively.The results demonstrate that the proposed method can achieve higher recognition accuracy and also has certain advantages in the face recognition under complex conditions.

3D face recognitiondeep learningdeep relationship perceptiondual domain feature enhancement

张龙、胡金蓉、张艳、黄果、黄飞虎

展开 >

成都信息工程大学 计算机学院,成都 610225

国家卫星气象中心,北京 100081

川西南空间效应探测与应用四川省高等学校重点实验室,四川 乐山 614004

3D人脸识别 深度学习 深度关系感知 双域特征增强

2025

电讯技术
中国西南电子技术研究所

电讯技术

北大核心
影响因子:0.472
ISSN:1001-893X
年,卷(期):2025.65(1)