Generalization-Enhanced Feature-Wise and Correlation Attentions for Cross-Database Facial Expression Recognition

Cross-database facial expression recognition (CDFER) has attracted increasing attention when evaluating the systems' generalization performance. Although the attention mechanism can capture the feature-wise importance or feature-correlation of expression sensitive regions, the attention-based network suffers from the overfitting to the source database, due to possible over-dependence on most salient features, without exploring feature characteristics during removal of feature redundancy. To address this issue, this paper introduces a multi-kernel competitive convolution in feature-wise attention to obtain more salient regions and let each kernel compete with others to enhance the expressive ability of features, by reducing attention overfitting to the source domain. For feature-correlation attention, we resort to a Monte Carlo-based dropout to not only reduce the over-learning of the feature relationship, but also model the dropout probability distribution more specifically by taking the characteristics of feature maps into account. Experimental results show that our algorithm achieves much better generalization performances than the state of the arts (SOTAs) on six publicly available datasets, in the scenarios of single source domain, multiple source domains and domain adaption.