Deep Learning for Facial Beauty Prediction: Integrating Transfer and Multi-Task Learning

Abstract

Facial beauty prediction is a complex and subjective task with significant applications in cosmetic surgery, virtual makeovers, social media filtering, and personalized beauty recommendations. The perception of beauty varies across individuals and cultures, making it challenging for computational models to generalize effectively. conventional approaches based upon handcrafted features and statistical models often fail to capture the intricate patterns of facial aesthetics, limiting their adaptability to diverse populations. To address these challenges, this study proposes an advanced facial beauty model that integrates transfer learning with multi-task learning for improving this predictive accuracy and generalization . The EfficientNetV2B0 architecture is employed to leverage its superior feature extraction capabilities and memory efficiency, whereas multi-task learning is utilized to jointly predict beauty scores alongside auxiliary tasks such as gender and ethnicity classification. This joint learning enables the model to learn shared representations and capture subtle aesthetic features more effectively. Experiments on the SCUT-FBP5500 benchmark dataset show that the proposed approach outperforms single-task learning models, achieving higher Pearson correlation coefficients and lower mean absolute errors. multi-task model incorporating beauty score, ethnicity and gender prediction achieve a Pearson correlation coefficient of 0.9185, mean absolute error of 0.2068, and root mean squared error of 0.2762, surpassing existing state-of-the-art methods. These findings suggest that integrating transfer learning with multi-task learning significantly enhances facial beauty prediction, facilitating for more robust and generalizable computational aesthetics models with real-world applications in automated beauty assessment.