Pose-robust recognition of low-resolution face images

Area of Science:

• Computer Vision
• Biometrics
• Machine Learning

Background:

• Surveillance face matching is hindered by low-resolution images, uncontrolled poses, and variable illumination.
• Existing methods struggle with the significant quality disparity between probe and gallery facial images.

Purpose of the Study:

• To develop an automatic approach for matching low-quality surveillance facial images with high-resolution enrollment images.
• To enhance the performance of face matching algorithms under adverse surveillance conditions.

Main Methods:

• Utilizing multidimensional scaling (MDS) to align feature spaces of probe and gallery images.
• Employing tensor analysis for facial landmark localization in low-resolution images.
• Developing a novel framework for simultaneous feature transformation.

Main Results:

• The proposed method effectively matches surveillance-quality faces to high-resolution images.
• Demonstrated superior performance compared to state-of-the-art super-resolution and classifier-based methods on the Multi-PIE dataset.
• Validated applicability on real-world surveillance imagery.

Conclusions:

• The developed approach significantly improves face matching accuracy for surveillance applications.
• The framework is effective for both tracking and recognition tasks in surveillance videos.
• Offers a robust solution for biometric identification in uncontrolled environments.