Periocular embedding learning with consistent knowledge distillation from face

Yoon Gyo Jung

doi:10.1016/j.neucom.2024.127263

Periocular embedding learning with consistent knowledge distillation from face

2024-03-01

0

PRE

AI

Leslie Ching Ow Tiong

A

Andrew Beng Jin Teoh *

DOI：10.1016/j.neucom.2024.127263

Original

Original request for help

Save

Abstract

En 中文

The periocular area, which refers to the peripheral area of the ocular, is a valid biometric for situations where facial recognition is not possible due to occlusion or masking. However, periocular biometrics alone can reduce discriminative information, particularly in wild environments. To address this, we propose Consistent Knowledge Distillation (CKD) that transfers discriminatory information from face images to periocular embeddings using temperature -based consistency. CKD achieves state-of-the-art results on challenging unconstrained periocular recognition benchmarks, improving performance by 49%-54% in relative gain. Furthermore, we provide insight into how CKD effectively extracts and transfers global inter -class relationship information by showing that CKD is equivalent to a learned -label smoothing approach with a novel sparsity -oriented regularizer.

Keywords:

Periocular recognition

Biometric identification

Knowledge distillation

Periocular embedding learning with consistent knowledge distillation from face

Abstract

Journal

Organization