融合多层级特征的跨年龄人脸识别方法  

Cross-age face recognition method incorporating multi-level features

作　　者：段文涛 智敏 DUAN Wen-tao;ZHI Min(College of Computer Science and Technology,Inner Mongolia Normal University,Hohhot 010000,China)

机构地区：[1]内蒙古师范大学计算机科学技术学院,内蒙古呼和浩特010000

出　　处：《计算机工程与设计》2024年第12期3732-3738,共7页Computer Engineering and Design

基　　金：内蒙古自然科学基金项目(2023MS06009、2018MS06008);内蒙古自治区高等学校科学研究基金项目(NJZZ21004)。

摘　　要：针对目前主流的跨年龄人脸识别方法中利用卷积神经网络的末层特征输出作为最终特征表示,忽视了卷积神经网络中间层次的特征表达,导致模型最后解耦到的身份特征不完整的问题,提出一种利用卷积神经网络中间层次的特征表达,包含特征选择融合模块和身份特征解耦模块的跨年龄人脸识别方法。基于注意力融合底层特征和高层语义获取混合特征;在多任务训练的监督下,非线性特征解耦获取身份特征;利用身份特征实现跨年龄人脸识别。该方法在人脸老化数据集AgeDB-30、CALFW和CACD-VS的准确率分别达到了96.89%、96.20%和99.60%,验证了其有效性。Feature representations within intermediate layers of convolutional neural networks are often overlooked by current mainstream cross-age face recognition methods,whereby the output of the final convolutional layer is solely utilized as the final feature representation.As a result,the issue of incomplete identity features in the decoupled model representation is encountered.A cross-age face recognition method was proposed,wherein feature representations from intermediate layers of convolutional neural networks were leveraged.The method was comprised of a feature selection fusion module and an identity feature decoupling module.A hybrid feature was obtained by fusing attention-based low-level features and high-level semantics.Under the supervision of multi-task training,identity features were extracted through non-linear feature decoupling.The cross-age face recognition was achieved by employing the extracted identity features.The effectiveness of the proposed method is demonstrated through accuracies of 96.89%,96.20%,and 99.60%achieved on the AgeDB-30,CALFW,and CACD-VS face aging datasets,respectively.

关 键 词：跨年龄人脸识别 特征融合 人脸识别 注意力机制 卷积神经网络 多任务训练 非线性特征解耦 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]