EFFICIENTNET FOR HUMAN FER USING TRANSFER LEARNING

Abstract

Automatic facial expression recognition (FER) remained a challenging problem in computer vision. Recognition of human facial expression is difficult for machine learning techniques since there is a variation in emotional expression from person to person. With the advancement in deep learning and the easy availability of digital data, this process has become more accessible. We proposed an efficient facial expression recognition model based EfficientNet as backbone architecture and trained the proposed model using the transfer learning technique. In this work, we have trained the network on publicly available emotion datasets (RAF-DB, FER-2013, CK+). We also used two ways to compare our trained model: inner and cross-data comparisons. In an internal comparison, the model achieved an accuracy of 81.68 % on DFEW and 71.02 % on FER-2013. In a cross-data comparison, the model trained on RAF-DB and tested on CK+ achieved 78.59%, while the model trained on RAF-DB and tested on FER-2013 achieved 56.10% accuracy. Finally, we generated an t-SEN distribution of our model on both datasets to demonstrate the model''s inter-class discriminatory power.