A two-phase approach for expression invariant 3D face recognition using fine-tuned VGG-16 and 3D-SIFT descriptors

Expression invariant 3D face recognition systems have many computer vision applications such as human-computer interaction. Most 3D face recognition systems rely on rigid region features and a substantial amount of training data to achieve better accuracy. However, the computational cost of these sy...

Full description

Saved in:
Bibliographic Details
Published inMultimedia tools and applications Vol. 82; no. 15; pp. 23873 - 23890
Main Authors P.R., Suganya Devi, R., Baskaran
Format Journal Article
LanguageEnglish
Published New York Springer US 01.06.2023
Springer Nature B.V
Subjects
Accuracy
Candidates
Computational efficiency
Computer Communication Networks
Computer Science
Computer vision
Computing costs
Data Structures and Information Theory
Face recognition
Facial recognition technology
Invariants
Methods
Multimedia
Multimedia Information Systems
Neural networks
Special Purpose and Application-Based Systems
Track 3: Biometrics and HCI
3D-SIFT
3D face recognition
Kernel principal component analysis
VGG-16
Face expression
Online AccessGet full text

Cover

More Information
Summary:Expression invariant 3D face recognition systems have many computer vision applications such as human-computer interaction. Most 3D face recognition systems rely on rigid region features and a substantial amount of training data to achieve better accuracy. However, the computational cost of these systems is very high. In order to address the problem of compromising computational efficiency for accuracy, this paper presents a computationally efficient two-phase expression invariant 3D face recognition approach using fine-tuned VGG-16 and 3D-SIFT descriptors. In the first phase, the pre-trained VGG-16 is fine-tuned with the Texas database and the CASIA-3D database. The candidates are recognized using the fused features from the fine-tuned VGG-16 and landmarks-based angles. In the second phase, the 3D-SIFT descriptors are computed on the rigid component of the probe and candidate 3D faces. Then, the final identity is obtained based on the best 3D-SIFT keypoints’ match score. Reporting competitive results in comparison to the state-of-the-art, the proposed method achieves 100% and 97.69% recognition rates respectively for the neutral-neutral and neutral-non neutral scenarios on the Bosphorus Database. Further, it requires only 1.27 seconds to identify a probe from a gallery with 105 faces.
ISSN:1380-7501
1573-7721
DOI:10.1007/s11042-023-14407-z