Graph Neural Networks Boost Kinship Verification Accuracy from Facial Images

A team of computer scientists has developed a graph neural network approach to kinship verification—the task of determining whether two people are biologically related by examining their facial features. The method achieves competitive results using individual face representations, matching the performance of earlier techniques that required joint analysis of both parties' faces.

The Challenge of Facial Kinship Detection#

Traditional kinship verification systems relied on extracting facial features from individual images and comparing them. However, these single-face approaches proved less reliable than methods that simultaneously analyzed parent-child facial pairs from the start, learning shared patterns between family members in one go.

A New Architecture with Graph Concepts#

The researchers introduced a system incorporating graph neural network principles to leverage standard face representations while achieving results comparable to joint-training methods. The approach includes a novel classification layer design and a blended loss function strategy that gradually introduces center loss during network training—a technique that encourages the model to group related faces closer together in its learned feature space.

Strong Experimental Results#

Testing on two standard datasets—KinFaceW-I and KinFaceW-II—demonstrated the method's effectiveness. On KinFaceW-II, the system achieved top-tier performance with an average gain of approximately 1.6 percentage points across all kinship relationship types. Performance on KinFaceW-I remained near the best-published results. The research team has made their implementation publicly available for other researchers to build upon.