Multimodal Attention Dynamic Fusion Network for Facial Micro-Expression Recognition

The emotional changes in facial micro-expressions are combinations of action units. researchers have revealed that units can be used as additional auxiliary data to improve micro-expression recognition. Most the attempt fuse image features and unit information. However, these works ignore impact on feature extraction process. Therefore, this paper proposes a local detail enhancement model based multimodal dynamic attention fusion network (MADFN) method for This uses masked autoencoder learnable class tokens remove areas with low expression ability images. Then, we utilize module representation potential features. state-of-the-art performance our proposed is evaluated verified SMIC, CASME II, SAMM, their combined 3DB-Combined datasets. experimental results demonstrated achieved competitive accuracy rates 81.71%, 82.11%, 77.21% SAMM datasets, respectively, show MADFN help discrimination