Adapting Vision Transformers with Recurrent Heads for Accurate Dynamic Emotion Recognition
Year: 2025
Type: conference-paper
Abstract: Dynamic Facial Emotion Recognition (DFER) is one of the fundamental challenges in computer vision, due to its computational complexity and the scarcity of large-scale video datasets. This study presents a novel and efficient approach for accurate recognition of dynamic emotions, applicable in areas such as healthcare, human–computer interaction, and behavioral analysis. The proposed work focuses on enhancing Vision Transformer (ViT)–based models through Parameter-Efficient Transfer Learning (PETL) to develop a precise yet lightweight architecture. To transfer knowledge from still images to videos, the basic Static to Dynamic (S2D) model is employed, and a proposed architecture with an advanced classification head based on an LSTM recurrent network and an attention mechanism is presented. This approach aims to achieve higher accuracy in emotion recognition by maintaining the complete structure of the basic model and enhancing the classification head to extract richer temporal information and weighting frames so that more significant frames contribute more to the final prediction. Evaluations on two benchmark datasets—Dynamic Facial Expressions in the Wild (DFEW) and Facial Expression Recognition Video dataset with 39K clips (FERV39k)—demonstrate that the proposed model achieves 75.35% Weighted Averaging Recall (WAR) on DFEW and 52.49% WAR on FERV39k, which is more accurate than existing methods.
Authors: Reza Ahmadi, Mahdi Hariri, Ali Amiri
Institution: University of Zanjan
Language: English
Topic: Emotion and Mood Recognition
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204207684
