Discriminative Region Transfer Network for Cross-Database Micro-Expression Recognition
Jianbang Li, Ruimin Hu, Mithun Mukherjee · 2022
Compared to conventional micro-expression analysis, cross-database micro-expression recognition (CDMER) considers a more practical scenario, where the training and testing samples come from different databases. Under this problem setting, most previous micro-expression recognition methods may suffer severe performance degradation due to dataset bias or domain shift. This makes CDMER a challenging but interesting problem. Most existing CDMER methods transfer global micro-expression features without considering the different contributions of facial regions to micro-expression. To tackle this problem, we first confirm the following argument: for samples of the same category from different datasets, their discriminative facial regions are similar and relevant. Therefore, we can transfer the knowledge about discriminative facial regions learned from the source domain to the target domain directly. Based on this argument, we design a novel deep domain adaption method called discriminative region transfer network (DRTN). The DRTN uses an adversarial-based adaptation method to align the source and the target distribution in the learned feature space. Then during transfer, we focus on the discriminative facial regions by the feat of attention mechanism. We conducted extensive experiments on CASME II and SMIC datasets and achieves very competitive results. These evaluations convincingly demonstrate the effectiveness of our proposed method.