The EfficientNet Performance for Facial Expressions Recognition

Convolutional neural networks (CNN) have advanced significantly in object detection, image classification, and various other implementions, including Facial Expression Recognition (FER). Modern CNN models become slower and require more computation as they become deeper and larger. In this context, EfficientNet is moving in a new path by using compound scaling to improve accuracy. In existing FER research, EfficientNet is used as a pre-trained network (as is or as a fine-tuned EfficientNet) and is then combined with a variety of classifiers. Previous studies indicate that the addition of the EfficientNet index does not result in an increase in accuracy EfficientNet was used in this study as feature extractor and followed by GAP, FCL, and Softmax Layer as classifier. The best model in this study was EfficientNetB3 (accuracy: 94.44%, F1 score: 0.94, and average ROC-curve score: 0.97). EfficientNet can recognize both positive and negative emotions. The variety of specific emotional derivatives and suppressed negative emotions pose difficulties in teacher emotion recognition. Although the research results show that EfficientNet outperforms other deep CNN models, it is requisite to reexamine the basic network and compound scaling formulation so that it can become a sophisticated model for the new Teachers FER dataset.