基于特征融合的轻量级新残差人脸识别方法  

Lightweight New Fesidual Face Recognition Method Based on Feature Fusion

作　　者：惠康华[1] 闫建青 高思华 贺怀清[1] HUI Kang-hua;YAN Jian-qing;GAO Si-hua;HE Huai-qing(College of Computer Science and Technology,Civil Aviation University of China,Tianjin 300300,China)

机构地区：[1]中国民航大学计算机科学与技术学院,天津300300

出　　处：《电子学报》2024年第3期937-944,共8页Acta Electronica Sinica

基　　金：国家重点研发计划项目(No.2020YFB1600101);天津市教委科研项目(No.2020KJ024)。

摘　　要：针对现有轻量级模型在嵌入式设备的人脸识别应用中存在识别精度难以提升的问题,提出一种融合人脸对齐关键特征点信息的轻量级新残差网络模型(Lightweight New Residual Network,LNRN).LNRN利用深度残差网络结构能够解决网络退化且避免干扰因素影响的优势,结合人脸对齐环节产生的关键特征点信息,对深度残差网络结构进行简化和合理设计,实现对关键特征信息和全局信息的提取.为避免特征提取过程中丢失重要特征信息,该模型在新残差网络中加入结合空间和通道的注意力机制进行辅助.在公开的四个标准人脸数据集上的仿真实验表明,该模型识别速度在接近主流轻量级人脸识别方法的同时,平均识别精度比MobiFace提高了0.6%.Aiming at the problem that the existing lightweight models are difficult to improve the recognition accura⁃cy in the face recognition applications of embedded devices,a new lightweight residual network model(Lightweight New Residual Network,LNRN)that integrated the key feature point information of face alignment is proposed.The advantage of deep residual network structure that can solve the network degradation and avoid the influence of interference factors are ab⁃sorbed by LNRN.In order to realize the extraction of key feature information and global information after combining the key point information generated by the face alignment,the deep residual network structure is simplified and reasonably de⁃signed.In order to avoid losing important feature information in the process of feature extraction,an attention mechanism combining space and channel is added to the new residual network for assistance.Simulation experiments on the four stan⁃dard face datasets showed that the recognition speed of the proposed model was close to the mainstream lightweight face methods,and the average recognition accuracy of the proposed model is 0.6%higher than that of MobiFace.

关 键 词：轻量级新残差网络模型 人脸识别 关键特征信息 注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]