Micro-expressions, fleeting and involuntary facial movements, are crucial for revealing concealed emotions but pose significant recognition challenges due to their subtlety and brief duration. Traditional methods, including handcrafted features and deep learning, often struggle with residual pose misalignments from imperfect facial alignment during preprocessing. This paper proposes a Facial Pose-Constrained Micro-Expression Recognition (MER) framework featuring a lightweight Orthogonal Bidirectional Long Short-Term Memory (O-BiLSTM) module. The module extracts fixed pose features from reference frames along orthogonal directions, providing spatial constraints to mitigate misalignment effects in the primary 3D ResNet network. Experiments on CASME II, SAMM, and SMIC datasets demonstrate state-of-the-art performance, with accuracies of 80.9%, 73.7%, and 70.1%, respectively. Ablation studies confirm the module’s efficacy in enhancing recognition by modeling global facial geometry. The method outperforms existing techniques, offering a robust solution for MER in applications like security and clinical diagnosis.