Improvement of facial attributes' estimation using Transfer Learning

Nowadays, we are experiencing the emergence of intelligent applications, capable of recognizing a human face by using shape, gender, face attributes or even emotions. Those application are deployed in numerous real-world sites, like facial recognition systems, Law enforcement applications or security purposes. For these reasons, we propose to improve available facial attributes' estimation in the main model, by adding other attributes, such as (medical-mask or face cover, head scarf, tattoo) and using transfer learning (TL). To do this end, using TL one stage, we suggest retraining the main model on new attributes all at once, and using TL multistage training, where we employ a TL network for each attribute. The main model is trained on the CelebA dataset with 40 attributes using a CNN model, while for the aforementioned three attributes, we use our constructed dataset. The obtained results show that the second method outruns the first in terms of metrics, but the first one is better in prediction rates, especially for the attributes of the main model, this is a problem caused by many TL networks losing data.