Enhancing Facial Expression Recognition Systems with a Stable Diffusion-Based Synthetic Fake Facial Emotion Recognition (FFER) Dataset

초록

Artificial intelligence (AI)-generated images are becoming challenging for people to discern from real-life ones because of recent advancements in synthetic data technology. This research aims to improve the accuracy of Facial Expression Recognition (FER) systems by introducing a deep learning-based classification model that uses synthetic images generated via a diffusion model. Initially, a synthetic dataset is constructed to replicate the seven classes present in existing FER datasets, providing a contrasting image set for comparison with real images. The study suggests combining a balanced dataset with our synthetic FFER dataset to classify facial expressions more accurately. Finally, the effectiveness of the proposed system is evaluated against alternative methodologies. The proposed system is compared to other works with different combinations of FER2013 dataset using FFER. The outcomes demonstrate the efficacy of the suggested techniques by showing that accuracy can be raised by 4% with FFER and 13% with a balanced dataset + FFER following the adoption of augmentation processes and generative models. The complete dataset formulated for this study, known as the FFER dataset, is publicly available to the research community for future investigations.

키워드