Diffusion-based learning for cross day hand gesture recognition using HD-sEMG signals

Gesture recognition using high-density surface electromyography (HD-sEMG) signals has attracted significant attention in myoelectric control. While recent studies report high intraday performance, interday accuracy often drops due to poor generalizability, limiting real-world deployment. To improve robustness, we propose a Diffusion-based Hand Gesture Recognition framework (DiffHGR) that integrates diffusion-based data augmentation with autoencoder representation learning. During training, a Diffusion (Diff) component corrupts HD-sEMG signals through a forward Gaussian diffusion process and employs a U-Net–based denoiser to reconstruct high-fidelity signals, which are used to augment the training set with diverse samples. Meanwhile, an Autoencoder (AE) component learns discriminative latent representations for gesture classification, enhanced via skip connections from the Diff encoder to reuse multi-scale denoising features. To address cross-day distribution shifts, we further introduce a lightweight few-shot calibration protocol. During calibration, the Diff is kept frozen and is used only as a generator to synthesize additional samples that augment the limited target-day data, while the AE encoder and classifier are updated for fast adaptation. During online inference, prediction is performed solely by the calibrated AE encoder and classifier, with the Diff generator inactive in the inference path, enabling low-latency deployment. Extensive experiments demonstrate that DiffHGR consistently outperforms other benchmark models. Real-time validation further confirms its robustness and practical applicability. These results highlight the effectiveness of combining diffusion-driven data augmentation and autoencoder-regularized representation learning for robust HD-sEMG-based gesture recognition.