Abstract

Facial beauty prediction (FBP) is a cutting-edge research area in artificial intelligence (AI), enabling computers to evaluate facial attractiveness similarly to human perception. Despite advancements in deep neural networks for FBP, challenges like limited label information and overfitting remain. Our study introduces a novel approach combining multi-task learning with an adaptive sharing policy and attentional feature fusion (AFF) to address these issues. Built on the AdaShare network with a ResNet18 backbone, the method enhances label utilization by leveraging multiple datasets and reduces overfitting through AFF’s attention mechanism, which integrates semantic information. This innovative framework significantly improves FBP accuracy by tackling insufficient label data and overfitting risks.  Experimental results on the Large-Scale Asia Facial Beauty Database (LSAFBD) and SCUT-FBP5500 datasets demonstrate superior performance compared to single-database, single-task baselines. The proposed method not only advances facial beauty prediction but also shows promise for broader applications in image classification and AI-driven tasks. This approach offers a robust solution for researchers seeking reliable and scalable AI models.