Simultaneous detection of face attributes via multi-task learning

Abstract

Identifying face attributes is an ongoing problem of research which is used in bio-metrics, surveillance etc. In past, researchers have proposed methods which predict single facial attribute at a time. Real-time applications need to predict multiple attributes simultaneously to increase their efficiency in real world. Recently, researchers have started to use Multi-Task learning to learn multiple facial attributes simultaneously and leverage the correlations among the tasks. Using MTL, related tasks can generalise better as compared to single task learning and it requires only one model to predict multiple features thus making it more efficient for testing in real time.In this research, we have proposed a Multi-Task learning architecture which simultaneously predicts gender, age and race of given input facial image. The model outperforms the methods which learn only single task at a time. We have used CelebA and UTKFace dataset to assess the effiectiveness of the proposed MTL architecture. We got best results for UTKFace dataset when we used SE-ResNet-50-128D as pretrained model with gender recognition accuracy equal to 98.47 %, race prediction accuracy equal to 87.56 % , and age Mean Squared Error equal to 4.68 years. We got best results for CelebA dataset when we used SE-ResNet-50-128D as pretrained model with gender recognition accuracy equal to 98.82 % and age prediction accuracy equal to 87.53 %. We compared these results against previous works on CelebA and UTKFace datasets. Proposed architecture outperformed current state-of-the-art architectures in most of the cases.