A lightweight attention-based network for micro-expression recognition

Micro-expression has emerged to be a feasible strategy in affective estimation due to its great reliability in emotion detection. Recent years have witnessed that deep learning methods were successfully applied to the micro-expression recognition field. In visual data, micro-expression only exists in regions such as eyebrows and mouth, etc, which leads to its imbalanced distribution. Therefore, it’s difficult for networks to distinguish the above micro-expression description with weak intensity when extracting feature maps. To tackle such issues, we propose a novel lightweight attention model, LAM, to improve the network recognition performance. LAM is enabled to calculate the correlation between the feature maps (channel dimension) and the correlation within the feature maps (spatial dimension), thus helping the network to focus on micro-expression information. Additionally, cooperated with residual block at various scales in Resnet, LAM can adaptively compute and update the feature maps in each network layer. Technically, coping with small datasets, we build LAM without adding obvious parameters, while a straightforward but efficient strategy that transfer facial expression knowledge is utilized together. Extensive experimental evaluations on two benchmarks (CASME II and SAMM) and post-hoc feature visualizations demonstrate the effectiveness of our proposed network with LAM.