Hierarchical Representation Learning for Kinship Verification.

Area of Science:

• Computer Vision
• Biometrics
• Machine Learning

Background:

• Kinship verification is crucial for image organization and human resemblance recognition.
• Understanding human perception of kinship cues is essential for developing automated systems.
• Existing methods require further enhancement for accurate kinship verification.

Purpose of the Study:

• To conduct a human study on kinship perception to identify key facial discriminators.
• To develop a novel deep learning framework (KVRL-fcDBN) for unsupervised representation learning of facial regions.
• To improve face verification accuracy by integrating kinship information as a soft biometric.

Main Methods:

• A human study analyzed participants' ability to recognize kin relationships from whole faces and specific facial regions.
• A hierarchical kinship verification via representation learning (KVRL) framework was employed.
• Filtered contractive deep belief networks (fcDBN) were proposed for feature representation, encoding relational information.
• A new WVU kinship database was created for research.

Main Results:

• The KVRL-fcDBN framework achieved state-of-the-art kinship verification accuracy on the WVU database and benchmark datasets.
• Kinship information, when used as a soft biometric, significantly boosted face verification performance.
• An improvement of over 20% in face verification was observed using the proposed KVRL-fcDBN framework.

Conclusions:

• The proposed KVRL-fcDBN deep learning framework offers a highly effective approach for kinship verification.
• Integrating kinship information as a soft biometric modality substantially enhances face verification systems.
• The study provides valuable insights into human kinship perception and its application in AI.