Prototype-Based Discriminative Feature Learning for Kinship Verification.

Area of Science:

• Computer Science, Artificial Intelligence, Machine Learning

Background:

• Traditional kinship verification relies on low-level hand-crafted features (e.g., LBP, Gabor).
• These features may not adequately capture complex familial relationships in facial images.

Purpose of the Study:

• To develop a novel method for learning discriminative mid-level features for kinship verification.
• To improve the characterization of kin relationships using learned features.

Main Methods:

• Proposed a prototype-based discriminative feature learning (PDFL) method.
• Constructed a reference set of unlabeled kin relation face samples.
• Represented training samples as mid-level feature vectors using Support Vector Machine (SVM) decision values.
• Formulated an optimization function to minimize intra-class distances and maximize inter-class distances.
• Introduced a multi-view PDFL approach to integrate multiple low-level features for enhanced mid-level feature learning.

Main Results:

• The proposed PDFL method demonstrated superior performance in kinship verification tasks.
• The multi-view PDFL further improved verification accuracy.
• Experimental results on four public datasets surpassed state-of-the-art methods and human performance.

Conclusions:

• The developed PDFL method effectively learns discriminative mid-level features for kinship verification.
• The multi-view extension enhances the robustness and accuracy of the kinship verification system.
• The proposed approach represents a significant advancement in automated kinship verification technology.