AdvKin: Adversarial Convolutional Network for Kinship Verification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Biometrics

Background:

• Kinship verification aims to identify familial relationships from facial images.
• Existing methods, including shallow learning and CNN-based approaches, struggle with recognizing subtle kinship cues.
• Family ID information and pairwise kin-face distribution differences are often overlooked.

Purpose of the Study:

• To propose a novel family ID-based adversarial convolutional network (AdvKin) for kinship verification.
• To enhance the recognition of discriminative kinship features in facial images.
• To address limitations in current kinship verification methods for both small-scale and large-scale datasets.

Main Methods:

• A self-adversarial mechanism using negative maximum mean discrepancy (NMMD) loss is employed in the first fully connected layer.
• Pairwise contrastive loss and family ID-based softmax loss are jointly used for supervised training in subsequent layers.
• A two-stream network architecture with residual connections forms the core of AdvKin.
• An ensemble of patch-wise AdvKin networks (E-AdvKin) is proposed for fine-grained feature augmentation.

Main Results:

• The proposed AdvKin model demonstrates superior performance compared to state-of-the-art approaches.
• Experiments were conducted on four small-scale KinFace datasets and the large-scale Families in the Wild (FIW) dataset.
• The method effectively utilizes family ID information and pairwise kin-face distributions.

Conclusions:

• AdvKin offers a significant advancement in kinship verification, particularly in challenging 'in the wild' scenarios.
• The proposed adversarial approach and feature augmentation techniques improve the accuracy of kinship recognition.
• The method shows promise for both research and practical applications in facial relationship identification.