کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

Dynamic Facial Expression Recognition using Vision Transform with GRU-Attention Temporal Head

نویسندگان :

Reza Ahmadi¹ Mahdi Hariri² Ali Amiri³

1- دانشگاه زنجان 2- دانشگاه زنجان 3- دانشگاه زنجان

کلمات کلیدی :

Facial Emotion Recognition،Parameter-Efficient،Vision Transformer (ViT)،Gated Recurrent Unit،Temporal Modeling

چکیده :

Dynamic Facial Expression Recognition presents significant challenges for edge computing and small-scale video datasets due to the computational demands and extensive parameters associated with state-of-the-art models. While Parameter-Efficient Transfer Learning techniques, such as Static-to-Dynamic (S2D), introduce adapter modules within Transformer blocks, they still involve a substantial number of trainable parameters (up to 9.26 million). In this study, we propose a non-intrusive architecture that prioritizes parameter efficiency for DFER. Our main idea is to use a Vision Transformer Base/16 (ViT-B/16), which is kept fully frozen after being pre-trained on a large static dataset. To model temporal dynamics and capture inter-frame dependencies, we add an ultra-lightweight head consisting of a GRU layer and a temporal attention mechanism to the end of the frozen ViT. This approach reduces the number of trainable parameters to just 2.12 million. Experimental results show that our proposed model achieves a 71.21% Weighted Average Recall (WAR) on the DFEW dataset while maintaining impressive computational efficiency, with a 77% reduction in parameters compared to S2D. Furthermore, it outperforms many heavier models with over 18 million parameters, establishing a new and optimal balance between accuracy and parameter efficiency for resource-constrained DFER systems.