Classification of Age Group, Gender, and Race from Facial Images Using Multi-Task Based Deep CNNs with Transfer Learning

This study evaluates Single Task Learning (STL) and Multi-Task Learning (MTL) approaches for classifying age group, gender, and race from facial images in the UTKFace dataset. STL-based VGG models achieved test accuracies of 92.2% (age group), 98.1% (gender), and 94.3% (race). MTL models were then developed using shared convolutional bases from VGG16, VGG19, ResNet50, and DenseNet121, combined with task-specific fully connected layers and transfer learning. Among these, MTL models with a VGG16 base consistently outperformed STL counterparts in both accuracy and F1-score. The best MTL model, using VGG16 with ImageNet weights, achieved accuracies of 92.5% (age group), 98.2% (gender), and 95.0% (race), while the same model with VGGFace weights yielded comparable results. Extended evaluations using 8-class and 10-class age group configurations demonstrated strong generalization, achieving age classification accuracies of 89.7% and 88.1% respectively, along with improved performance in gender and race classifications. Partially frozen MTL models further reduced parameter counts while maintaining strong performance. These findings demonstrate the effectiveness of MTL in enhancing classification performance and computational efficiency for multi-attribute facial analysis.