Lightweight Models in Face Attribute Recognition: Performance Under Oclussions

In this paper we will study the performance of several light-weight convolutional neural networks with respect to state-of-the-art models for facial attribute classification. Specifically, we will try to determine the attributes of gender, age and ethnicity. There are many models based on lightweight architectures, from which we have chosen MobileNet and EfficientNet. The results obtained match or improve the state of the art in gender and race, achieving good results in age classification as well. On the other hand, we have performed a comparative study of these classifications with respect to two datasets. The first dataset is UTK-Face which contains the facial images aligned and a higher number of individuals, having a lower total number of samples, while the second dataset is VGG-Face2 which has a much higher total number of samples, having fewer individuals than UTK-Face and with a lower quality facial alignment.