Maximal likelihood correspondence estimation for face recognition across pose.

Area of Science:

• Computer Vision
• Biometrics
• Machine Learning

Background:

• Conventional face recognition methods struggle with pose variations due to feature misalignment.
• Existing image matching methods for cross-pose recognition have limitations in exploiting facial structure and personalization.

Purpose of the Study:

• To address limitations in existing cross-pose face recognition methods.
• To develop a robust method for estimating semantic correspondence between faces in different poses.
• To enable personalized correspondence learning for each probe image.

Main Methods:

• Developed a Morphable Displacement Field (MDF) model to encode face-specific structure information for semantic correspondence.
• Proposed a Maximal Likelihood Correspondence Estimation (MLCE) method for personalized correspondence learning.
• Synthesized virtual frontal face images from profile views for subsequent recognition using linear discriminant analysis with pixel-intensity features.

Main Results:

• Achieved state-of-the-art performance on multipose benchmarks (CMU-PIE, FERET, MultiPIE).
• Demonstrated reliable correspondence learning in complex, real-world environments (Labeled Face in the Wild database).
• The MDF regularization and MLCE objective proved effective for robust cross-pose face recognition.

Conclusions:

• The proposed MDF and MLCE methods significantly improve face recognition accuracy across varying poses.
• The approach effectively handles complex environmental conditions and diverse facial poses.
• This work advances the field of cross-pose face recognition by enabling personalized and structurally informed correspondence estimation.