<p>Facial micro-expression recognition remains challenging because extremely subtle, spatially localized facial muscle activations are easily diluted by globally dominant feature representations, resulting in insufficient regional discrimination. To address this issue, we propose a novel Multi-scale Local Detail Enhancement (MSLDE) model that jointly performs discriminative region localization and subtle feature learning through a recurrent global–local collaborative learning. Specifically, the global emotional features extracted by the Cycle Generative Adversarial Classification Network (CycleGACN) are used to guide an Attention Localization Network (ALN) to predict informative regions of interest (ROIs). The localized ROIs are then resized and fed into parallel convolutional branches to learn fine-grained local representations. During training, global classification loss, local classification losses, global–local ranking loss, and adversarial loss are jointly optimized, enabling accurate ROI prediction and facilitating subtle feature extraction. In contrast, refined local predictions provide feedback to improve global feature representation and ROI ranking. We conducted a comprehensive experiment, and the results showed that the MSLDE had achieved superb performance on the three micro-expression datasets of SMIC, CASME II, SAMM, and their composite 3DB-Combined dataset.</p>