Enhancing Face Emotion Recognition with FACS-Based Synthetic Dataset Using Deep Learning Models

In our study, we propose an innovative approach for facial emotion recognition that employs both transfer learning and deep learning techniques. Our methodology allows automatic detection of visual cues linked to various facial emotions. The approach is highly consistent and effectively addresses the challenge of detecting seven fundamental human emotions, which include anger, disgust, fear, happiness, sadness, neutral, and surprise. Additionally, we introduce a new dataset named EMOTE-2023, which is developed using Unreal Engine and the Maya platform. Also, the proposed approach is analyzed on other existing datasets, such as FER2013 and CK+. To enhance the accuracy of our emotion classification, we have used multiple deep learning models VGG, ResNet, MobileNet to verify the effectiveness of our Facial Emotion Recognition system. Our proposed approach achieved the highest accuracy of 72.49%, 96.97%, and 97.97% using the ResNet18 model for FER2013, CK+ and EMOTE-2023 respectively. Our work involves leveraging transfer learning techniques that use pre-trained neural network layers and trained on larger datasets that are commonly employed in computer vision applications.