Research for Facial Beauty Prediction Combined Multi-Task Transfer Learning with Knowledge Distillation

 At present, there are some problems in facial beauty prediction, such as few data samples, unclear evaluation index and large change of face appearance. Multi-task transfer learning can effectively utilize the additional useful information of related tasks and source domain tasks. Knowledge Distillation can distill some knowledge of teacher model into student model, and reduce the complexity and size of model. In this paper, multi-task transfer learning and knowledge distillation are combined for facial beauty prediction, in which facial beauty prediction using Large Scale Asian Facial Beauty Database (LSAFBD) is the main task and gender identification in SCUT-FBP5500 database is regarded as the auxiliary task. Firstly, multi-input multi-task facial beauty teacher model and student model are constructed. Secondly, we trained the multi-task teacher model and calculated its soft targets. Finally, knowledge distillation is carried out by combining the soft targets of the multi-task teacher model and the soft and hard targets of the student model. Experimental results show that the multi-task teacher model achieves an accuracy of 68.23% in the facial beauty prediction, which has a complicated structure and a parameter of 14793K. Although the multi-task student model achieves an accuracy of 67.39% after knowledge distillation, its structure is simple and parameter is only 1366K. The classification accuracy of the multi-task teacher model is higher than that of other methods. Although the classification accuracy of the multi-task student model is slightly lower, the model is simpler and the network parameters are less. It is more advantageous to use the lighter network to predict the facial beauty.