NAS4FBP: Facial Beauty Prediction Based on Neural Architecture Search

Abstract

AbstractFacial Beauty Prediction (FBP) is an important task in image processing, which simulates human perception of facial beauty. In related studies, most methods are based on canonical convolutional backbones. However, can the canonical backbones perform best in FBP? To tackle this problem, we propose a NAS4FBP framework, which adopts a multi-task neural architecture search strategy to auto determine the backbone structure. In our multi-task learning scheme, we propose HBLoss to better reveal the nature of facial aesthetic hierarchy. In addition, we introduce a new pre-processing method to enhance the data diversity and propose a non-local spatial attention module, to further improve the model performance. Our model achieves 0.9387 PC on the SCUT-FBP5500 benchmark dataset, surpassing other related models and reaching a new state-of-the-art.KeywordsFacial beauty predictionNeural architecture searchMulti-task learning