基于双重注意力机制的深度人脸表示算法

针对现有模型很少对人脸特征进行设计且人脸特征区分性较弱的问题, 提出一种基于双重注意力机制的深度人脸表示算法. 该算法采用双重注意力机制的网络结构, 通过细节注意力机制设计低层特征, 自动和自适应地学习层次特征, 关注局部特征; 通过语义注意力机制设计高层特征, 自适应地进行语义分组, 关注语义特征. 在LFW,YTF,MegaFace,IJB-B和IJB-C数据集上的实验结果表明, 双重注意力机制方法的识别精确度分别高达99.87%,97.9%,98.91%,95.02%和96.28%, 比同类算法Groupface平均提升了0.02%,0.1%,0.2%,1%和1%, 表明了双重注意力机制网络的优势.

Deep Face Representation Algorithm Based on Dual Attention Mechanism

Aiming at the problem that the existing models rarely designed face features and the face features were weak in discrimination, we proposed a deep face representation algorithm based on dual attention mechanism. The algorithm adopted network structure of dual attention mechanism, designed low-level features through detail attention mechanism, and paid attention to local features through automatic and adaptive learning of hierarchical features, semantic attention mechanism was used to design high-level features, and pay attention to semantic features through adaptive semantic grouping. The experimental results on LFW,YTF,MegaFace,IJB-B and IJB-C datasets show that the recognition accuracy of the dual attention mechanism method is as high as 99.87%,97.9%,98.91%,95.02% and 96.28% respectively, which is 0.02%,0.1%,0.2%,1% and 1% higher than that of similar algorithm Groupface. The comparative experiments show  the advantages of dual attention mechanism network.