Facial deblur inference using subspace analysis for recognition of blurred faces

Area of Science:

• Computer Vision
• Image Processing
• Biometrics

Background:

• Face recognition systems struggle with blurred images.
• Inferring the blur characteristics (Point Spread Function - PSF) from a single image is an ill-posed problem.

Purpose of the Study:

• To develop a novel method for face recognition that effectively handles blurred facial images.
• To accurately infer the Point Spread Function (PSF) of blur in facial images for subsequent deblurring.

Main Methods:

• Constructing a feature space where faces blurred by the same PSF are clustered.
• Learning statistical models of predefined PSF sets within this feature space.
• Matching query images to the closest PSF model for inference and deblurring.

Main Results:

• Substantial improvement in face recognition performance on artificially blurred datasets (FERET) compared to existing methods.
• Demonstrated enhanced performance on real-world blurred images from the FRGC 1.0 database.
• Showcased further performance gains by integrating facial deblur inference with the Local Phase Quantization (LPQ) method.

Conclusions:

• The proposed method effectively addresses the challenge of recognizing blurred faces by accurately inferring and correcting blur.
• Learned prior information and a specialized feature space significantly improve PSF inference tractability.
• The approach offers a robust solution for face recognition in unconstrained environments with varying blur conditions.