Deep convolutional multi-informative metric correlation analysis with bottleneck attention module for face recognition in the wild

Deep convolutional multi-informative metric correlation analysis with bottleneck attention module for face recognition in the wild

연구 분야: Strategies

학회: Multimedia Tools and Applications

Feature representation and verification have always been a difficult problem in the classification performance of Automatic Face Recognition in Unconstrained Conditions (AFR-UC). It is very important to select informative features to train a classifier, which is prerequisite. Inspired by the great success of Attention Mechanism (AM), we address the problem of face target recognition by proposing a novel feature representation method which takes advantage of exploiting the extracted deep features from Bottleneck Attention Modules (BAMs) on face images to introduce more powerful informative features and robust representation ability. First, the proposed Face Oriented Net is fine-tuned on eight face datasets including: ORL, Faces94, Grimace, Jaffe, Asian, Hispanic, Black and Multiracial. Second, the face oriented network is used for extracting deep features from input data using two BAM modules. Third, the extracted deep features are fused by using a traditional summation and a Multiple-Discriminant Informative Analysis (MICA) algorithm. Finally, based on LogDet Divergence-Based Metric Learning Triplet Constraints (LDMLT), K-Nearest Neighbors algorithm (K-NN) is proposed for face classification. Experiments on the eight face datasets are conducted and the classification accuracy results demonstrate that the proposed method outperforms the state-of-the-art methods.