摘要
针对口罩遮挡下注意力机制存在的多维度动态协同能力不足与细粒度抑制欠缺等问题,本文提出一种基于双重注意力校准的鲁棒识别方法,在通道和空间两个维度进行遮挡区域的动态校准。其中通道维度基于全局统计量抑制污染通道的异常响应,空间维度则定位遮挡区域并削弱其梯度传播,实现了从粗粒度筛选到细粒度增强的动态校准。在此基础之上,通过加权交叉熵损失和三元组损失进一步引导模型聚焦了局部未遮挡区域的特征表达,从而扩大类间特征距离间隔。实验结果表明,本文提出的双重注意力校准机制经过通道维度特征筛选与空间维度区域增强的协同作用,本文算法在LFW与AgeDB-30的掩膜场景下,相比于ArcFace算法,分别提高了6%和7.2%的准确率,在真实遮挡数据集MAFA数据集上则提高了7.3%,验证了其在复杂遮挡场景下的识别鲁棒性。
This paper proposes a robust recognition method based on dual attention calibration to address the issues of insufficient multi-dimensional dynamic collaboration and fine-grained suppression in the attention mechanism under mask occlusion.The method dynamically calibrates the occlusion area in both channel and spatial dimensions.The channel dimension is based on global statistics to suppress abnormal responses of polluted channels,while the spatial dimension locates occluded areas and weakens their gradient propagation,achieving dynamic calibration from coarse-grained screening to fine-grained enhancement.On this basis,the weighted cross entropy loss and triplet loss are used to further guide the model to focus on the feature expression of locally unobstructed areas,thereby expanding the inter class feature distance interval.The experimental results show that the dual attention calibration mechanism proposed in this paper,through the synergistic effect of channel dimension feature screening and spatial dimension region enhancement,has improved accuracy by 6%and 7.2%respectively compared to the ArcFace algorithm in mask scenes of LFW and AgeDB-30,and by 7.3%on the real occlusion dataset MAFA dataset,verifying its recognition robustness in complex occlusion scenes.
作者
傅鹏飞
徐巍
刘怀广
刘源泂
刘劲威
Fu Pengfei;Xu Wei;Liu Huaiguang;Liu Yuanjiong;Liu Jinwei(School of Mechanical Automation,Wuhan University of Science and Technology,Wuhan 430081,China;Key Laboratory of Metallurgical Equipment and its Control,Ministry of Education,Wuhan University of Science and Technology,Wuhan 430081,China;Hubei Provincial Key Laboratory of Mechanical Transmission and Manufacturing Engineering,Wuhan 430081,China)
出处
《电子测量技术》
北大核心
2025年第22期177-186,共10页
Electronic Measurement Technology
关键词
口罩遮挡人脸识别
注意力机制
双重注意力校准
加权交叉熵损失
三元组损失
mask covering facial recognition
attention mechanism
dual attention calibration
weighted cross entropy loss
triplet loss
