Multi-scale Temporal Attention Network for Blepharospasm Recognition in Videos

Facial movement disorders are prevalent in the field of medicine, with blepharospasm being a prototypical disorder affecting the muscles surrounding the eye. The integration of artificial intelligence technology for fine-grained recognition and digital-assisted diagnosis of blepharospasm holds significant implications. This study presents an analysis and investigation of blepharospasm based on video data. A 3D CNN-based multi-scale temporal attention network (MTA-Net) is proposed, which possesses the capability to accurately localize spatial regions and represent facial movement instances at various visual tempos. In order to equip the MTA-Net with global spatial-temporal features processing ability, we propose a Spatial Separable Coordinate Attention (SSCA) module for effectively modeling global spatial facial features, as well as an MTA module for capturing multi-scale global temporal motion features. The significant performance enhancements of MTA-Net are achieved through the complementary integration of our SSCA and MTA modules. The sensitivity of the proposed model is $97.21 \%$, which is higher than that of other existing models.