Human Age Group Classification Using CNN and Vision Transformer-Based Feature Extraction

Abstract

The human face reveals significant information about an individual’s identity, age, gender, emotion, and ethnicity. In face-to-face communication, age plays a vital role, influencing perception and interaction. However, facial features change over time due to aging, including alterations in skin thickness, color, texture, subcutaneous tissue composition, and the development of wrinkles and skeletal structure changes. These variations, which differ widely among individuals due to intrinsic and extrinsic factors, make accurate age estimation from facial images a challenging task. To address this problem, we propose a robust age estimation model that combines Vision Transformers (ViT) and Convolutional Neural Networks (CNN) for feature extraction. We utilize three classifiers Support Vector Machines (SVM), CNN with SoftMax, and k-Nearest Neighbors (KNN) for age group classification. Our research focuses specifically on Ethiopian individuals aged 15 to 60, using a dataset of 8,100 facial images collected from the Ethiopian Immigration Office and various other organizations. Each of the nine age classes includes 900 images. Our experimental results show that the CNN model achieved an accuracy of 91.4%, ViT achieved 97.8%, and the combined CNN+ViT model reached 98.5% accuracy using the SVM classifier with PCA. The ensemble approach outperformed individual models, demonstrating the effectiveness of integrating CNN and ViT for age estimation. This work highlights the potential of deep learning-based ensemble methods in improving age estimation accuracy and has promising applications in areas such as Human-Computer Interaction (HCI), surveillance, web content filtering, and electronic customer relationship management (e-CRM).