Comparison and Analysis of CNN Models to Improve a Facial Emotion Classification Accuracy for Koreans and East Asians

Abstract

Facial emotion recognition is one of the popular tasks in computer vision.  Face recognition techniques based on deep learning can provide the best face recognition performance, but using these techniques requires a lot of labeled face data. Available large-scale facial datasets are predominantly Western and contain very few Asians. We found that models trained using these datasets were less accurate at identifying Asians than Westerners. Therefore, to increase the accuracy of Asians' facial identification, we compared and analyzed various CNN models that had been previously studied. We also added Asian faces and face data in realistic situations to the existing dataset and compared the results. As a result of model comparison, VGG16 and Xception models showed high prediction rates for facial emotion recognition in this study. and the more diverse the dataset, the higher the prediction rate. The prediction rate of the East Asian dataset for the model trained on FER2013 was relatively low. However, for data learned with KFE, the model prediction of FER2013 was predicted to be relatively high. However, because the number of East Asian datasets is small, caution is needed in interpretation. Through this study, it was confirmed that large CNN models can be used for facial emotion analysis, but that selection of an appropriate model is essential. In addition, it was confirmed once again that a variety of datasets and the prediction rate increase as a large amount of data is learned.