ETwins: Enhanced Twins Transformer for Facial Expression Recognition

Facial expression recognition (FER) requires distinguishing between subtle visual differences, making it a challenging fine-grained classification task. This paper presents ETwins, an Enhanced Twins transformer architecture specifically tailored for facial expression recognition. ETwins builds on the Twins vision transformer and incorporates three key improvements: a weighted global average pooling layer that emphasizes critical facial regions, the integration of a class token for capturing both local and global features, and an enhanced positional encoding mechanism to embed essential contextual information. Our evaluations on widely used FER datasets demonstrate that ETwins achieves improved accuracy compared to larger baseline models while using fewer parameters. This result underscores the efficiency of custom vision transformers for fine-grained tasks, achieving competitive FER performance with a compact, parameter-efficient design.

کلیدواژه ها:

Facial Expression Recognition ، Vision Transformers ، Fine-Grained Image Classification ، Deep Learning