A Bayesian Scene-Prior-Based Deep Network Model for Face Verification.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Face recognition/verification is crucial but challenged by limited training samples.
• Deep learning excels but often requires large, representative datasets.
• Existing methods suffer performance drops with insufficient data per individual.

Purpose of the Study:

• To address the performance degradation of deep learning face recognition with few training samples.
• To propose a novel method that utilizes background scene information to enhance face feature extraction.
• To improve the robustness and accuracy of face verification systems under data scarcity.

Main Methods:

• Developed a Bayesian scene-prior-based deep learning model.
• Extracted scene semantics from diverse backgrounds to generate new face scenarios.
• Employed scene domain transfer to create augmented face images with varied backgrounds.

Main Results:

• Achieved 99.2% accuracy on the Labeled Faces in the Wild (LFW) dataset (view #2).
• Reached 94.3% accuracy on the YouTube Faces database.
• Demonstrated significant performance improvement with reduced training samples compared to literature benchmarks.

Conclusions:

• Background scene information is vital for robust face feature representation.
• The proposed Bayesian scene-prior model effectively enhances face verification accuracy, particularly in low-data regimes.
• Scene transfer learning offers a promising direction for improving deep learning-based face recognition systems.