Two-Stage Meta-Learning with Matched Feature Regularization for Cross-Subject sEMG Gesture Recognition Under Posture Variation.

Surface electromyography (sEMG)-based gesture recognition has attracted considerable attention in intelligent prosthesis control, human-computer interaction, and rehabilitation assistance. However, practical deployment remains challenging because new users can usually provide only a few labeled samples for calibration. Under cross-subject and posture-varying conditions, this setting further aggravates distribution shifts and can destabilize target-domain adaptation. To address these issues, this paper proposes a cross-subject sEMG gesture-recognition framework integrating training-time augmentation, two-stage meta-transfer learning, and matched feature regularization. Model-agnostic meta-learning (MAML) is first used on source-subject data to learn an initialization for rapid transfer, after which target-subject fine-tuning is performed with an auxiliary class-consistent feature constraint. Experiments were conducted on a self-collected 11-subject multi-posture dataset using a leave-one-subject-out (LOSO) protocol under FT-1_full, which uses one complete calibration repetition, and FT-2_k10, which uses 10 windows per gesture per posture. The proposed MAML + FT + MATCHED method achieved higher average Accuracy and Macro-F1 than Source-pretrain + FT. Under FT-2_k10, the average improvements were 9.20 and 9.70 percentage points, respectively. Nevertheless, the primary subject-level paired Wilcoxon comparisons did not reach statistical significance (Accuracy: p = 0.2402; Macro-F1: p = 0.2402; Holm-adjusted p = 1.0000 for both). The results therefore indicate favorable average trends and positive effect sizes, while pairwise statistical superiority was not established in the present 11-subject cohort.