Publication

Abstract : Due to the broad application potential and business demand, facial expression recognition is one of the prevalent topics in the deep learning domain. Since most of the application domain of facial expression is deployed on devices with low computational power, this work majorly focuses on a lightweight model. Therefore, in the proposed model, a lightweight vision transformer model, MobileVit that surpasses other light weight CNN like MobileNetV3, NasNet on the ImageNet classification dataset is chosen for the classification of facial expression. The model was initially pretrained on CelebA dataset, then was trained on RAF-DB dataset, and fine-tuning of the model was carried out. For improving the generalization capability of the model, advanced augmentation methods like cut mix with basic augmentation methods are deployed. For the visualization of the model performance, interpretation of feature maps was carried out utilizing the Score-Cam technique. Squeeze-and-excitation layer was added to improve the feature extraction performance of the model. After fine-tuning of the model, the performance of the model was evaluated on KDEF, FER-2013, and AffectNEt dataset.