<p>Micro-expression is subtle facial expressions that individuals inadvertently reveal, which is difficult to recognize owing to its shorter duration and lower intensity compared to traditional expressions. Current state-of-the-art algorithms based on video-based deep learning methods have achieved impressive recognition performance by training on video datasets captured under well-lit conditions. However, the performance of recognition generally suffers from poor robustness and generalization capabilities in under-illuminated environments. To address the issue, we propose a low-light facial micro-expression recognition (LL-FMER) model based on Transformer. Firstly, a progressive low-light noise augmentation (PLNA) training strategy is proposed to enable the network to gradually adapt and learn low-light features of inputs without altering the training dataset. Besides, a low-light optical-flow prediction estimator (LOPE) is introduced to accurately predict optical flow under low-light conditions. Finally, the Transformer-based self-attention network enhances the accuracy of motion feature extraction by capturing both local and global features across hierarchical structures. Extensive experiments indicate that the proposed algorithm surpasses various state-of-the-art approaches under both normal-light and low-light conditions.</p>