Low Resolution Face Recognition Across Variations in Pose and Illumination

Area of Science:

• Computer Vision
• Biometrics
• Machine Learning

Background:

• Recognizing low-resolution face images in uncontrolled environments presents significant challenges.
• Existing methods often struggle with variations in image quality and capture conditions.

Purpose of the Study:

• To develop a fully automatic approach for robust face recognition from low-resolution images.
• To improve the accuracy and efficiency of face recognition systems operating in real-world scenarios.

Main Methods:

• Utilizing multidimensional scaling (MDS) to learn a common transformation matrix for aligning low- and high-resolution face images.
• Employing stereo matching cost to quantify image similarity in the transformed space.
• Introducing a reference-based approach to optimize computational efficiency by using stereo matching costs from select reference images.

Main Results:

• The proposed method demonstrates effective recognition performance on challenging, real-world face image databases.
• The reference-based approach significantly reduces computational time while maintaining high accuracy.
• Experimental results show superiority compared to state-of-the-art super-resolution and cross-modal synthesis techniques.

Conclusions:

• The developed automatic approach offers a promising solution for low-resolution face recognition in uncontrolled environments.
• The integration of MDS and a reference-based stereo matching cost provides an efficient and accurate face recognition system.
• This algorithm effectively addresses the limitations of current face recognition technologies in handling degraded image quality.