Cross-Cultural Analysis of Attention Regions in Facial Expression Recognition Model

​

This study evaluates the facial expression recognition performance of
deep learning models using Japanese and Swedish facial expression data sets, and
in the process, to elucidate cross-cultural differences in feature extraction of the
models through visualization and principal component analysis. Specifically, a
convolutional neural network based on VGG16 is trained on Japanese and
Swedish facial expression data, and intra-culture evaluation and cross-culture
evaluation are compared. We use LayerCAM to visualize which regions of the
image the trained model focuses on strongly, and explore how the model's focus
changes when the culture is different. Principal component analysis is used to
numerically capture the distribution of regions of focus for each emotion to better
understand cross-cultural effect and feature differences by emotion. The results
of this study provide insights into the importance of dataset and model design
that takes cultural context into account in making facial expression recognition
systems more versatile.