基于强化学习的模块化有遮挡人脸识别算法研究  

Research on modular obscured face recognition algorithm based on reinforcement learning

作　　者：王丹淋 张辅超 刘昕 王丽丽 WANG Danlin;ZHANG Fuchao;LIU Xin;WANG Lili(State Grid Materials Co.,Ltd.,Beijing 100120,China;Liaoning Planning and Designing Institute of Post and Telecommunication Co.,Ltd.,Shenyang 110179,China)

机构地区：[1]国网物资有限公司,北京100120 [2]辽宁邮电规划设计院有限公司,辽宁沈阳110179

出　　处：《电子设计工程》2025年第5期157-161,共5页Electronic Design Engineering

摘　　要：针对当前遮挡人脸识别算法存在准确率偏低及鲁棒性较差的不足,文中基于分块稀疏和强化学习模块提出了一种有遮挡的人脸识别算法。对于传统稀疏算法全局特征提取能力强、局部特征提取能力弱的问题,该算法结合分块的思想构建了分块稀疏模型,从而提升了模型的识别准确率。强化学习模块使用MobileNet作为源网络,并采用通道与空间注意力机制来增强网络的自适应性能。同时利用剪枝网络对模型的卷积核进行简化,从而增强了其与嵌入式移动设备间的兼容性。实验结果表明,所提算法的识别准确率在不同数据集上相较于对比算法提高了约5.01%和4.81%,参数量减少约0.267 MB,FPS则提升约12 fps,充分证明了该算法的鲁棒性及计算效率均较为理想。Aiming at the shortcomings of low accuracy and poor robustness of current occluded face recognition algorithms,this paper proposes an occluded face recognition algorithm based on block sparse and reinforcement learning module.For the problem of strong global feature extraction ability and poor local feature extraction ability of traditional sparse algorithm,the new algorithm constructs a block sparse model by combining the idea of blocking,thus improving the accuracy of model recognition.The reinforcement learning module uses MobileNet as the source network,and adopts channel attention and spatial attention mechanisms to improve the network adaptive performance.At the same time,the pruning network is used to simplify the model convolution core,which enhances the compatibility with embedded mobile devices.The experimental results show that the recognition accuracy of the proposed algorithm is about 5.01%and 4.81%higher than that of the comparison algorithm in different data sets,the parameter amount is reduced by about 0.267 MB,and the FPS is improved by about 12 fps,which fully proves that the robustness and computational efficiency of the algorithm are ideal.

关 键 词：稀疏表示 分块思想 MobileNet-v1 注意力机制 剪枝网络 人脸识别 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TN60[自动化与计算机技术—计算机科学与技术]