AM3F-FlowNet: Attention-Based Multi-Scale Multi-Branch Flow Network

Micro-expressions are the small, brief facial expression changes that humans momentarily show during emotional experiences, and their data annotation is complicated, which leads to scarcity of micro-expression data. To extract salient distinguishing features from a limited dataset, we propose an attention-based multi-scale, multi-modal, multi-branch flow network thoroughly learn motion information micro-expressions by exploiting attention mechanism complementary properties between different optical information. First, (horizontal flow, vertical strain) based on onset apex frames videos, each branch learns one kind separately. Second, multi-scale fusion module more prosperous stable feature expressions using spatial focus locally important at scale. Then, design multi-optical reweighting adaptively select for separately channel attention. Finally, better integrate three branches alleviate problem uneven distribution samples, introduce logarithmically adjusted prior knowledge weighting loss. This loss function weights prediction scores samples categories mitigate negative impact category imbalance classification process. The effectiveness proposed model demonstrated through extensive experiments visualization benchmark datasets (CASMEII, SAMM, SMIC), its performance comparable state-of-the-art methods.