Adaptively Weighted k-Tuple Metric Network for Kinship Verification

Area of Science:

• Computer Vision
• Biometrics
• Artificial Intelligence

Background:

• Facial image-based kinship verification is crucial in biometrics.
• Existing methods often overlook discriminative features from multiple negative pairs, limiting generalizability.
• Human visual systems utilize both low-order and high-order cross-pair information.

Purpose of the Study:

• To develop a novel deep learning model for enhanced facial kinship verification.
• To address the limitations of existing approaches by incorporating high-order cross-pair features.
• To improve the generalizability and performance of kinship verification systems.

Main Methods:

• Proposed the adaptively weighted k-tuple metric network (AW k-TMN), an end-to-end deep learning model.
• Introduced a novel cross-pair metric learning loss based on k-tuplet loss to capture low-order and high-order features.
• Implemented an adaptively weighted scheme to emphasize hard negative examples and utilized multi-level convolutional features.

Main Results:

• The AW k-TMN model demonstrated superior performance compared to state-of-the-art approaches.
• Experiments on three popular kinship verification datasets validated the model's effectiveness.
• The approach successfully leveraged high-order cross-pair features for improved accuracy.

Conclusions:

• The proposed AW k-TMN effectively enhances facial kinship verification by utilizing comprehensive feature information.
• The novel loss function and adaptive weighting scheme contribute to improved discriminative power.
• The study provides a robust deep learning solution for kinship verification with released code and models.