Neighborhood repulsed metric learning for kinship verification

Area of Science:

• Computer Vision
• Machine Learning
• Biometrics

Background:

• Facial kinship verification is a challenging computer vision task with limited existing research.
• High similarity between non-related individuals can lead to misclassification.

Purpose of the Study:

• To develop a novel metric learning method for improved facial kinship verification.
• To enhance discrimination between related and unrelated individuals using facial images.

Main Methods:

• Proposed a Neighborhood Repulsed Metric Learning (NRML) approach.
• Developed a Multiview NRML (MNRML) method for fusing multiple feature descriptors.
• Learned a distance metric to pull related individuals closer and push unrelated individuals further apart.

Main Results:

• The proposed NRML and MNRML methods demonstrated significant efficacy in kinship verification.
• Multiview NRML improved performance by effectively fusing complementary feature information.
• Experimental results showed the proposed methods are comparable to human observer performance.

Conclusions:

• The NRML and MNRML methods offer a promising approach for automated facial kinship verification.
• The developed techniques effectively leverage discriminative information for accurate verification.
• The study validates the potential of machine learning for complex biometric tasks.