Towards Open-Set Myoelectric Gesture Recognition via Dual-Perspective Inconsistency Learning

TL;DR

This work tackles open-set recognition for sEMG-based gesture understanding by revealing a pronounced prediction inconsistency of unknowns across dual perspectives. It introduces PredIN, a two-branch prototype-learning framework that maximizes cross-branch class-feature distribution inconsistency via an inconsistency loss while preserving per-branch discriminability with a triplet loss. The approach achieves state-of-the-art open-set rejection and maintains strong closed-set accuracy across multiple public sEMG datasets, demonstrating robust OSR performance in real-world HMI scenarios. The results suggest dual-perspective inconsistency learning as a practical and effective strategy for open-set recognition in biomedical signal processing and may extend to other pattern recognition domains.

Abstract

Gesture recognition based on surface electromyography (sEMG) has achieved significant progress in human-machine interaction (HMI), especially in prosthetic control and movement rehabilitation. However, accurately recognizing predefined gestures within a closed set is still inadequate in practice; a robust open-set system needs to effectively reject unknown gestures while correctly classifying known ones, which is rarely explored in the field of myoelectric gesture recognition. To handle this challenge, we first report a significant distinction in prediction inconsistency discovered for unknown classes, which arises from different perspectives and can substantially enhance open-set recognition performance. Based on this insight, we propose a novel dual-perspective inconsistency learning approach, PredIN, to explicitly magnify the prediction inconsistency by enhancing the inconsistency of class feature distribution within different perspectives. Specifically, PredIN maximizes the class feature distribution inconsistency among the dual perspectives to enhance their differences. Meanwhile, it optimizes inter-class separability within an individual perspective to maintain individual performance. Comprehensive experiments on various benchmark datasets demonstrate that the PredIN outperforms state-of-the-art methods by a clear margin. Our proposed method simultaneously achieves accurate closed-set classification for predefined gestures and effective rejection for unknown gestures, exhibiting its efficacy and superiority in open-set gesture recognition based on sEMG.

Figures (15)

• Figure 1: An illustration of open-set myoelectric gesture recognition system. An open-set myoelectric gesture recognition system need to correctly classify predefined known gestures while rejecting unknown gestures to avoid generating false interaction signals.
• Figure 3: An illustration of our proposed framework. Our framework consists of dual perspectives represented by two branches, each of which contains an encoder and a set of learnable prototypes. $\mathcal{L}_{PL}$ and $\mathcal{L}_{trip}$ are applied to each branch individually while $\mathcal{L}_{incon}$ simultaneously acts on both. $\mathcal{L}_{incon}$ aims to maximize the class feature distribution inconsistency, ensuring each branch has an entirely distinct layout or neighboring class pairs (blue arrows). Upon this, unknown samples like $\mathbf{z}_{u}$ represented by purple are predicted near the clusters of different known classes due to prediction inconsistency while known samples agree on the same predictions across two perspectives.
• Figure 4: An illustration of class feature distribution inconsistency between dual perspectives. Different colors represent different classes.
• Figure 5: An illustration of how inconsistency loss $\mathcal{L}_{incon}$ acts on two branches to adjust the class feature distribution. The class pairs of each branch are optimized in opposite directions. When the class pairs are pulled within the margin, the optimization of $\mathcal{L}_{incon}$ will halt for one branch.
• Figure 6: Rejection rules. Through averaging, unknown samples (right) tend to predict different results and produce lower $S_{max}$, while known samples (left) tend to obtain the same predictions consistent with the label (yellow) and produce higher $S_{max}$.