Attribute-induced Attractiveness Regression of Facial Images with Multi-task Convolution Neural Network

Facial attractiveness prediction is a significant problem that involves many research areas. Recently, deep-learning-based methods have made significant progress for this task, while many challenges remain tackled. Firstly, it is hard to guarantee the objectivity of manual evaluation facial attractiveness, which results in labeling bias of training data. Secondly, the training data with attractiveness score required for a robust predictor is in severe shortage. For the former problem, we adopt the ELO rating algorithm prevalently used in sports to dynamically reorder the ranking of training images according to the comparisons between images. We propose a method based on attribute-induced regression with a multi-task Convolution Neural Network to overcome training data shortage for the last challenge. The method employs facial attributes of facial parts such as eyes, nose, mouth to restrain and reduce the facial attractiveness prediction problem’s solution space. We collected, labeled, and ranked an Asian female face dataset with attributes for facial attractiveness analysis. Comprehensive experiments conducted on the dataset demonstrate the effectiveness of our method.