Multiclass Regression for Facial Beauty Prediction Based on Deep Learning Using SCUT-B 5500

Abstract

FBP, that is, facial beauty prediction, is a fundamental procedure of how beautiful a face person perceives, just like human beings. The challenge focuses on systems that can assess facial features and provide ratings that align with human perceptions of attractiveness. In this paper, we investigate the usage of deep learning techniques using ResNet18 models for predicting beauty of a face using SCUT-B 5500 dataset and share our findings. In the last ten years machine recognition and scoring of attractiveness has developed into a new field through the use of artificial intelligence. We present our exploratory research on constructing a robust model based on a dataset containing 5500 annotated frontal images ranked according to perceived beauty. Multi-task transfer learning was employed to improve the model performance and address the issue of limited data. Our ResNet18 model had an impressive accuracy of over 91% on predicting beauty ratings. Furthermore, this study not only contributes to the field of facial beauty prediction, but it also has the potential to be implemented in multiple fields such as social networks, dating applications, personalized ads.