Face Attribute Analysis Method Based on Self-Supervised Siamese Network

Abstract

Face attribute analysis, which is a challenging and popular task in the vision field, has been widely used in various fields including intelligent security, human-computer interaction, and targeted promotion. However, in practical applications, people are not always facing the camera, and their head postures would affect the accuracy of attribute prediction. In addition, some attributes are inherently more difficult to predict due to the image noise, motion blur, and light variation. To improve the accuracy of face attribute analysis, a self-supervised Siamese multi-task convolutional neural network (SS-MCNN) is proposed in this paper. First, a Siamese network model is built for multi-task joint training. Second, a sparse self-supervised loss function is designed to learn the common features of Siamese contrastive data. Finally, the proposed SS-MCNN improves the performance on 40 face attributes analysis with an average accuracy of 91.42% on CelebA dataset. For age estimation task, the model also achieves good results with a mean absolute error (MAE) of 3.29 and 3.10 on the AFAD and MORPH II test sets.