Transfer Learning-Based Ethnicity Recognition Using Arbitrary Images Captured Through Diverse Imaging Sensors.

Area of Science:

• Computer Science
• Artificial Intelligence
• Biometrics

Background:

• Ethnicity recognition is crucial for various applications but faces challenges like facial feature variations and class imbalance.
• Machine learning models for ethnicity classification require robust and accurate predictive capabilities.

Purpose of the Study:

• To introduce a novel experimental framework for ethnicity recognition.
• To leverage transfer learning and optimize Convolutional Neural Network (CNN) models for improved classification performance.
• To evaluate the impact of different imaging sensors and real-world conditions on facial recognition accuracy.

Main Methods:

• Utilized transfer learning with pre-trained CNN models (VGG16, DenseNet169, VGG19, ResNet50, MobileNetV2, InceptionV3, EfficientNetB4) on the UTKFace and a custom celebrity dataset.
• Performed rigorous preprocessing and hyperparameter optimization to identify an ideal configuration for optimal performance.
• Examined the influence of various imaging sensors and real-world capture conditions on model performance.

Main Results:

• The VGG19 model achieved 87% validation accuracy and 75% maximum test accuracy on the celebrity dataset.
• The optimized VGG19 model demonstrated an overall accuracy of 87% across five ethnic groups, with per-class accuracies ranging from 51% to over 90%.
• Comparative experiments confirmed competitive performance against recent state-of-the-art methods.

Conclusions:

• Transfer learning on sensor-captured images enables robust and efficient ethnicity classification.
• Systematic hyperparameter optimization enhances model generalization and mitigates overfitting.
• The proposed framework offers a reproducible and fairness-aware approach to ethnicity recognition.