Supervised Mixed Norm Autoencoder for Kinship Verification in Unconstrained Videos.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Kinship identification is crucial for organizing vast online video content.
• Current kinship verification methods primarily rely on image pairs, limiting video analysis.
• Unconstrained videos present challenges like varying illumination, pose, and occlusion.

Purpose of the Study:

• To propose a novel deep learning framework for kinship verification in unconstrained videos.
• To introduce a Supervised Mixed Norm regularization Autoencoder (SMNAE) for enhanced feature learning.
• To develop a robust system capable of handling diverse video conditions.

Main Methods:

• A three-stage deep learning framework utilizing a novel Supervised Mixed Norm regularization Autoencoder (SMNAE).
• SMNAE incorporates class-specific sparsity in its weight matrix for improved representation.
• The framework leverages learned spatio-temporal features from video frames for verification.

Main Results:

• The proposed SMNAE framework achieved 83.18% accuracy on the newly collected KIVI database.
• This represents a performance improvement of at least 3.2% compared to existing algorithms.
• Consistent superior results were observed across the KIVI database and six other public kinship databases.

Conclusions:

• The SMNAE-based deep learning framework offers a significant advancement in video-based kinship verification.
• The novel autoencoder formulation effectively captures essential spatio-temporal information for kinship identification.
• The framework demonstrates strong generalizability and effectiveness across multiple diverse datasets.