Enhancing Face Recognition Model Performance Through HyperStyle‐Driven Data Augmentation
Applied Computational Intelligence and Soft Computing · 2025-01 · 2025 (1)
DOI: 10.1155/acis/4097213 ↗Kurz gefasst
Fachartikel in der Zeitschrift Applied Computational Intelligence and Soft Computing, online erschienen am 17.02.2025. Verfasst von Muhammad Chaidir und 3 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 44 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Universitas Syiah Kuala.
Abstract (Originalsprache)
Face recognition is a reliable biometric technology that utilizes artificial intelligence and computer vision to identify and verify an individual’s identity based on facial features. Its rapid development has enabled integration into diverse applications, including security systems, access control, forensics, and mobile devices. However, challenges such as recognizing faces across different ages persist, presenting opportunities for further research. Developing robust face recognition models requires substantial training datasets, but manually labeling large collections of facial images is both time‐intensive and costly. Our previous study demonstrated that a ResNet‐50 model trained on a dataset of 44,000 facial images, combining the original FaceScrub dataset with HyperStyle age and smile augmentations, achieved an F1‐score of 79%, significantly outperforming the model trained solely on the authentic FaceScrub dataset (F1‐score: 63%). Building upon these findings, this study explores a broader range of augmentation styles, including pose and domain adaptation, e.g., Pixar, Toonify, Sketch, and Disney, using the HyperStyle scheme. Models trained with domain adaptation showed reduced performance (F1‐score: 64%) due to the domain gap introduced by cartoon‐like styles. In contrast, models trained with pose augmentation achieved an F1‐score of 75%, demonstrating the importance of augmentations resembling real‐world variations. When training incorporated all augmentation styles, e.g., age, smile, pose, and domain adaptation, the F1‐score reached 78%, slightly lower than the model trained with age and smile augmentations only (F1‐score: 79%). To further evaluate the impact of age and smile augmentations, additional experiments were conducted using various CNN models, including VGGNet‐16, MobileNetV3Small, SEResNet‐18, and ResNet‐50, under different hyperparameter configurations. Among these, ResNet‐50 achieved the highest F1‐score (82%), surpassing VGGNet‐16 (60%), MobileNetV3Small (65%), and SEResNet‐18 (81%). The study also introduced modifications to the HyperStyle scheme in both the preprocessing and postprocessing stages, which further enhanced model performance. The ResNet‐50 model trained with the modified HyperStyle scheme, and the original FaceScrub dataset achieved the highest F1‐score of 83%, demonstrating the effectiveness of these modifications in improving face recognition performance.
Förderung
- Universitas Syiah Kuala
Zitieren
APA
Chaidir, M., Abidin, T. F., Sofyan, H., & Muchtar, K. (2025). Enhancing Face Recognition Model Performance Through HyperStyle‐Driven Data Augmentation. Applied Computational Intelligence and Soft Computing, 2025(1). https://doi.org/10.1155/acis/4097213
Vancouver
Chaidir M, Abidin TF, Sofyan H, Muchtar K. Enhancing Face Recognition Model Performance Through HyperStyle‐Driven Data Augmentation. Applied Computational Intelligence and Soft Computing. 2025;2025(1). doi:10.1155/acis/4097213
BibTeX
@article{Chaidir2025,
author = {Chaidir, Muhammad and Abidin, Taufik F. and Sofyan, Hizir and Muchtar, Kahlil},
title = {Enhancing Face Recognition Model Performance Through HyperStyle‐Driven Data Augmentation},
journal = {Applied Computational Intelligence and Soft Computing},
year = {2025},
volume = {2025},
number = {1},
doi = {10.1155/acis/4097213}
}
EUVIMED