Authors:
Fakunle Ajewole、Joseph Damilola Akinyemi、Khadijat Tope Ladoja、Olufade Falade Williams Onifade
Paper:
https://arxiv.org/abs/2408.06806
Introduction
Biometrics, the computerized identification of individuals based on genetic and behavioral characteristics, has become increasingly reliable and accurate. Among various biometric tools, face recognition stands out due to its high accuracy, non-contact nature, user-friendliness, speed, and dependability. Age-Invariant Face Recognition (AIFR) is crucial in many applications, including access control, distribution of government benefits, and criminal investigations, as it ensures the identification of individuals despite age differences between enrolled and probe images.
Despite significant research efforts in AIFR, African ethnicity has been underrepresented or misrepresented, often conflating indigenous Africans with African-Americans. This study aims to address this gap by developing an AIFR system specifically for indigenous African faces, using a dataset of 5,000 images from 500 individuals across 10 African countries.
Literature Review
Research on AIFR has employed various machine learning techniques, datasets, and methods. Key studies include:
- 3D Ageing Modelling: Improved performance on FG-NET, MORPH, and BROWNS databases.
- Maximum Entropy Feature Descriptor (MEFD): Achieved 76.2% and 92.26% accuracy on FG-NET and MORPH, respectively.
- Deep Convolutional Neural Networks (DCNN): Decomposed deep facial characteristics into age-related and identity-related aspects.
- Application Invariant Model (AIM): Provided a unified deep architecture for age-invariant facial representations.
- Mutual Information Minimization (MT-MIM): Disentangled face features into identity-dependent and age-dependent components, achieving significant improvements on FG-NET, MORPH, CACD, and AgeDB datasets.
Despite these advancements, there has been limited attention to indigenous African ethnicity in AIFR research, highlighting the need for this study.
Methodology
The AIFR problem was formulated as a multiclass closed-set recognition problem. The study used two datasets: FAGE v2, a newly curated dataset of indigenous African faces, and a subset of the CACD dataset representing African-Americans.
Dataset Collection
- FAGE v2: Collected images of 500 individuals from 10 African countries, evenly distributed across five geographical zones. Each individual had 10 age-separated images, totaling 5,000 images.
- CACD Subset: Extracted images of African-American actors, resulting in 8,656 images of 89 individuals.
Model Architecture
The VGGFace architecture was employed for feature extraction and classification. The model was fine-tuned with varying hyperparameters, including the number of training epochs, batch size, number of dense layers, and dropout fraction.
Experimental Setup
Each dataset was split into training, validation, and test sets. The model’s performance was evaluated using classification accuracy, precision, recall, and F1-score.
Experiments, Results, and Discussions
Five experiments were conducted with different hyperparameter combinations. The best accuracies were obtained with 50 training epochs, a batch size of 64, one fully connected layer, and a 50% dropout rate.
Recognition Accuracies
- FAGE v2: Best accuracy of 81.8%
- CACD: Best accuracy of 91.5%
Performance Evaluation
The performance metrics for both datasets are presented in Tables 4 and 5.
Comparative Analysis
A comparative analysis was conducted by selecting 10 images from each of the 88 African-American individuals in CACD and 88 individuals from FAGE, resulting in 880 images each. The results revealed that CACD images were better predicted than FAGE v2 images, indicating differences in the aging features of indigenous versus non-indigenous Africans.
Conclusion
This study developed an AIFR model for indigenous African faces, achieving a recognition accuracy of 81.8%. The findings highlight significant differences in the aging patterns of indigenous and non-indigenous Africans, suggesting the need for further exploration and development of robust face recognition algorithms for African ethnicities. Future work will focus on enlarging the FAGE v2 dataset and investigating larger subsets of non-indigenous Africans from different datasets.
