Deep Face Rectification for 360° Dual-Fisheye Cameras

Area of Science:

• Computer Vision
• Biometrics
• Image Processing

Background:

• Rectilinear face recognition models perform poorly on fisheye images.
• 360° back-to-back dual fisheye cameras capture highly distorted images.
• Fisheye image distortion severely impacts face recognition accuracy.

Purpose of the Study:

• To propose a novel face rectification method for fisheye images.
• To improve the performance of face recognition systems using fisheye imagery.
• To address the challenges posed by non-linear fisheye projection in biometrics.

Main Methods:

• Developed a two-part network: a classification network for distortion level and a restoration network for geometric correction.
• Integrated the rectification method into a conventional rectilinear face recognition system.
• Tested the end-to-end system on synthetic (Labeled Faces in the Wild) and real-world fisheye datasets.

Main Results:

• Achieved 99.18% face verification accuracy on synthetic LFW dataset.
• Attained 95.70% face verification accuracy on a real image dataset.
• Demonstrated an average accuracy improvement of 6.57% for verification and 4.51% for identification compared to conventional systems.

Conclusions:

• The proposed face rectification method effectively mitigates fisheye distortion.
• The integrated system shows significant performance gains in face recognition tasks.
• This approach enhances the applicability of face recognition in scenarios using fisheye cameras.