Face image analysis via Transformers: Algorithms and

The perception of beauty has long been a central topic in human society, shaped
by socioeconomic, cultural, and historical influences. Despite the evolving opinions
on facial beauty worldwide, understanding the factors behind facial attractiveness
remains a key area of interest across disciplines such as psychology, computer
science, and evolutionary biology. With advancements in computer vision and
deep learning, facial beauty prediction (FBP) has emerged as a cutting-edge field,
enabling objective quantification of facial beauty and its underlying factors.
This thesis proposes four novel approaches to facial beauty prediction using
deep learning. Two approaches leverage convolutional neural networks (CNNs)
integrated with ensemble learning, combining predictions from multiple models
to improve accuracy. The remaining two approaches harness the power of Vision
Transformers, utilizing attention mechanisms to capture intricate relationships
within facial features. Together, these methods enhance feature representation and
analysis for robust and reliable facial beauty assessment.
Experiments conducted on the SCUT-FBP5500 benchmark dataset demonstrate
the effectiveness of our approaches. We achieve superior performance by comparing
various deep learning models, including AlexNet, ResNet-18, and ResNeXt-50. The
proposed models yield predictions that closely align with human evaluations, surpassing
conventional methods in accuracy and consistency. This thesis underscores
the transformative impact of deep learning in facial beauty prediction, offering
precise, unbiased, and automated evaluations of facial attractiveness.