Improving cross-resolution face matching using ensemble-based co-transfer learning.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biometrics

Background:

• Face recognition systems typically perform poorly when matching low-resolution images (e.g., from surveillance) against high-resolution databases.
• Extracting discriminative features from low-resolution faces is challenging, especially with limited labeled training data.

Purpose of the Study:

• To address the challenge of cross-resolution face matching, specifically matching low-resolution probe images with high-resolution gallery images.
• To propose and evaluate a novel co-transfer learning framework for this task.

Main Methods:

• A co-transfer learning framework combining transfer learning and co-training paradigms is proposed.
• Transfer learning adapts knowledge from high-resolution matching to low-resolution probes.
• Co-training assigns pseudo-labels to unlabeled low-resolution data to facilitate knowledge transfer.

Main Results:

• The proposed co-transfer learning framework significantly enhances cross-resolution face recognition performance.
• Experiments on multiple face databases demonstrate the algorithm's efficacy compared to existing methods and a commercial system.
• The approach proves useful in real-world scenarios with challenging conditions.

Conclusions:

• The amalgamation of transfer learning and co-training in an ensemble framework effectively improves cross-resolution face matching.
• The proposed method offers a robust solution for matching low-resolution faces to high-resolution galleries, addressing a critical gap in biometrics.