Infant Facial Expression Recognition Based on Parameter-free Attention Module

Abstract

In order to further improve the accuracy of infant facial expression recognition, an infant facial expression dataset was established, and a parameter-free attention module (PFAM) was proposed. Firstly, the images about infants were collected through the Internet. After screening, 17785 images were chosen and divided into five categories, namely happiness, sadness, surprised, sleeping, and neutral, which generally reflect the infant facial expression. Secondly, using the average pooling and max pooling characteristics in the feature map channel and space, we proposed the parameter-free attention module. Finally, the recognition rate was compared to the common attention module and the deeper residual network. The experimental results show that the recognition rate of Resnet18 network with the PFAM is superior to attention modules SE and CBAM and deeper residual networks, and the recognition rate on self-built infant facial expression dataset exceeds that of the ResNetl01, and the recognition rate on public facial expression dataset RAF-DB exceeds that of the ResNetl52.