MicroSwin-ESBO: A Swin Transformer Optimized Framework for Facial Micro-Expression Detection and Classification

Facial micro-expressions (MEs) are fleeting, involuntary facial movements that reveal suppressed or hidden emotions and typically last less than half a second, making their detection and classification a challenging task. To address this, we propose a novel framework called Micro-Expression Swin Transformer with Enhanced Secretary Bird Optimization (MicroSwin- ESBO) for Facial Micro-Emotion Detection and Classification (FMEDC), which integrates robust pre-processing, shallow feature learning, advanced deep modelling, and metaheuristic optimization. The pre-processing pipeline includes face detection and alignment using Distributed Library (Dlib) and Retina-based Face Detector (RetinaFace), Radial Transformation (RT) for precise registration, and Facial Cropping (FC) based on landmark distances, followed by motion modelling with the Total Variation-L1 Optical Flow (TV-L1 OF) algorithm and Optical Flow Strain (OFS) calculation to highlight subtle muscle dynamics. For shallow feature extraction, a Spatio-Temporal Strain Tensor Network (STSTNet)-inspired multi-branch architecture is employed to capture essential motion cues while suppressing redundant information, and these features are further processed using the Shifted Window Transformer (Swin Transformer, Swin-T) backbone to capture both local and global spatiotemporal dependencies. To optimize classification performance, the Enhanced Secretary Bird Optimization (ESBO) algorithm is applied for hyperparameter tuning, and finally, a Softmax Classifier (SC) categorizes the micro-expressions into distinct emotional states. The proposed MicroSwin-ESBO framework provides a reliable and efficient solution for Micro- Expression Recognition (MER), with promising applications in criminology, Human–Computer Interaction (HCI), Online Learning (OL), and Psychological Assessment (PA).