Sep 2026· IEEE Transactions on Mobile Computing· Vol 25, pp. 15280-15292· 0 citations· 76 references
Abstract
The implementation of surface electromyography (sEMG)-based hand gesture recognition on mobile and wearable systems is frequently restricted by the finite computing, memory, and battery capabilities of edge devices. Even though a low-density sEMG setup is a feasible hardware implementation, achieving robust recognition under such constraint conditions becomes very challenging due to the non-stationary nature and inter-subject variance. In this paper, we propose DSCAttenEMG, an efficient neural network that combines both Depthwise Separable Convolution (DSC) for local feature extraction and Multi-Head Self-Attention (MHSA) to model long-range dependencies on EMG/IMU data, using 1× 1 DSC followed by Global Average Pooling to replace high-dimensional fully connected layers. Extensive experimentation on a self-collected dataset, the public SeNic and BandMyo datasets shows that our approach achieves state-of-the-art recognition performance (94.45%, 94.11% and 92.89%) at negligible complexity (only 178–179 K parameters). The model is capable of real-time inference (0.93 ms on RTX 4090 GPU, 6.68 ms on NVIDIA Jetson AGX Orin, 1.4/0.7 ms on CPU/NPU of Qualcomm mobile platform) and has a high degree of practicality for embedded deployment (118 samples/s at <inline-formula><tex-math notation="LaTeX">$\sim$</tex-math><alternatives><mml:math><mml:mo>∼</mml:mo></mml:math><inline-graphic xlink:href="wen-ieq1-3697898.gif"/></alternatives></inline-formula>1 W on K230 edge AI platform). This amalgamation of three pivotal strengths, elevated accuracy, enhanced efficiency, and pragmatic viability, highlights its substantial potential for practical mobile and wearable applications.
Objective. Gesture recognition using surface electromyography (sEMG) is a promising technology for wearable human–machine interaction systems. However, sEMG-based gesture classification models often suffer from performance degradation in cross-user scenarios due to the variability in individual physiological signals. To address this issue, large-scale deep learning algorithms with a great amount of parameters are employed to enhance model generalization and improve recognition performance across different users. However, the use of deep learning models on wearable devices is challenging due to their limited computational power and memory capacity. Methods. This study addresses this challenge through knowledge distillation, which compresses large teacher models into lightweight student models while maintaining performance. We trained deep teacher models, including DenseNet, InceptionV3, and VggNet, and distilled their parameters into compact convolutional neural network and long short-term memory (LSTM) models. The experiment was performed on the sEMG data of 30 subjects collected from a wrist-band electrode with 32 channels. Main Results. The results demonstrated that student models, particularly LSTM-based ones, achieved classification accuracy close to or even higher than their teacher models, with a highest accuracy improvement. Among the teacher–student combinations, the DenseNet121-LSTM architecture achieved the highest classification accuracy. The relationship between computational complexity floating-point operations (FLOPs) and model performance was also analyzed, showing that the distilled models can effectively approximate high-FLOP models. Significance. This research enhances the feasibility of embedding deep learning models into wearable systems (e.g. wristband or wristwatch), facilitating more responsive and efficient gesture recognition in daily-life applications. The Python implementation of the complete knowledge distillation framework is publicly available at: https://github.com/Open-EXG/handDistill.
Fang Qiu, Chenyun Dai, Xiaodong Liu et al.· Journal of Neural Engineerin...· 0 citations
Surface electromyography (sEMG) is increasingly used for gesture recognition in prosthetics, rehabilitation, and human–computer interaction. Existing architectures typically force heterogeneous sEMG features into a shared latent representation, limiting their ability to capture complementary temporal, frequency-domain, and inter-electrode spatial dependencies. To better exploit these features, this paper proposes a three-branch fusion network. Unlike many existing multi-branch methods, the proposed network explicitly models the ring arrangement of armband electrodes, capturing the adjacency information in the sensor topology that linear channel representations ignore. The temporal and spectral branches use a compact multi-scale residual structure, so this topology branch is added while maintaining modest model complexity. A reliability-aware routing mechanism then adaptively assigns fusion weights to the three branches for each sample. On NinaPro DB5 Exercise B (eight-channel lower armband), the method reaches 83.99% under subject-dependent training and 85.66% under transfer learning, exceeding prior transfer learning approaches under matched conditions. Ablation experiments confirm that the three branches contribute non-redundant information and that adaptive fusion outperforms fixed combinations. The architecture also generalizes to MyoArmbandDataset under a subject-adaptive transfer learning protocol without dataset-specific hyperparameter retuning, indicating potential for wearable gesture interfaces, rehabilitation, and prosthetic control.
Luoqi Cui, Yong Liu, Hadi Fathollahi Abdar et al.· Italian National Conference...· 0 citations
Accurate hand gesture recognition using surface electromyography (sEMG) typically relies on multichannel sensor arrays and computationally intensive models. This limits practical deployment in low-power and embedded systems. This study investigates the feasibility of classifying ten hand gestures using a single sEMG channel combined with lightweight machine learning architectures. Raw sEMG signals were transformed into a comprehensive feature-based representation, including time-domain, frequency-domain, higher-order-crossing, and relative-intensity features. Feature redundancy was reduced using Pearson correlation filtering and the removal of highly correlated features, while dimensionality-reduction techniques (LDA and PCA) were applied selectively. Three classifiers, a feed-forward neural network (NN), k-nearest neighbors (KNN), and a support vector machine (SVM), were systematically evaluated across four experiments. Results demonstrate that combining time and frequency features with Pearson filtering and a compact NN can achieve up to 90 percent accuracy, even with limited temporal and spatial information. These findings highlight the potential of single-channel sEMG systems for cost-effective, low-power gesture-recognition applications.
Hand gesture recognition via surface electromyography (sEMG) is fundamental to prosthetic control. In this field, deep learning approaches have become the gold standard. However, current architectures struggle to scale; model performance typically decreases as the number of hand movements increases. Performance degradation is tied to the increased statistical complexity of decoding expanded gesture sets and compounded by the limitations of state-of-the-art methods, which primarily rely on low-latency unimodal convolutional architectures. Convolutions operate locally, limiting model's ability to capture long-range sequential patterns. Unimodal setups cannot leverage complementary information from coordinated signals characterizing movement execution, such as inertial and eye-tracking data. These limitations motivate architectures that integrate local and global features across multimodal physiological sequences. To bridge this gap, this study introduces EMG-CrossFormer, an end-to-end hybrid convolutional-transformer for seamless multimodal integration. EMG-CrossFormer combines representations from an arbitrary number of unimodal encoders through cascaded cross-attention fusion layers, and decodes the fused representations using learnable gesture queries. EMG-CrossFormer was evaluated on four NinaPro datasets (DB2, DB3, DB7, and DB10) and benchmarked against six state-of-the-art models using an increasing number of modalities. Using only sEMG, EMG-CrossFormer achieved mean accuracies of 72.33%, 52.48%, 79.16%, and 73.49% on DB2, DB3, DB7, and DB10, respectively. Incorporating inertial signals improved performance to 90.66%, 80.40%, 92.79%, and 92.06%. These results show that joint local-global feature modeling improves sEMG-only decoding and that multimodal fusion substantially amplifies this benefit, underscoring the value of both design principles for complex hand gesture recognition.
Federico Del Pup, E. Tentori, M. Atzori· 0 citations
Electromyography (EMG) serves as a robust signal source for developing intuitive human-machine interfaces (HMIs). With the proliferation of commercial and prototype EMG armbands, there is a growing need to balance high-accuracy gesture recognition with hardware constraints. While deep learning architectures have achieved remarkable performance, they often require significant computational resources and high-end sensors. This study investigates an algorithmic-centric perspective, exploring whether sophisticated feature extraction (FE) can compensate for reduced hardware specifications. Using a 20-gesture dataset collected from 10 subjects, we evaluated performance across three hardware configurations: 500 Hz/8-bit (C1), 500 Hz/12-bit (C2), and 1000 Hz/8-bit (C3). We compared traditional FE methods against modern paradigms designed to capture spatial and spatio-temporal signal dynamics, specifically PHASOR, Myoelectric Temporal Patching (MTP), and the proposed WaveLSTM (Wave long short-term memory) framework. WaveLSTM embeds deep-learning-inspired processing schemes, i.e., utilizing short- and long-term memory components, into a computationally efficient pipeline with significantly lower overhead than standard deep learning models. Our results demonstrate that WaveLSTM achieves superior accuracy, exceeding 95% in a leave-one-trial-out validation scheme. Notably, while classical FE approaches exhibited lower overall recognition rates, WaveLSTM demonstrated hardware invariance, maintaining high performance regardless of changes in sampling frequency (p=0.275) or ADC resolution (p=0.941). These findings suggest that advanced spatio-temporal feature engineering can enable high-performance, complex gesture recognition on low-cost, low-power wearable devices without sacrificing accuracy.
Rami N. Khushaba, F. Verdini, A. Mengarelli et al.· 2026 IEEE/ASME International...· 0 citations
: Due to the accelerating aging of the population and the continuous increase in the number of stroke patients, there is an urgent need for rehabilitation robotics technology with high-precision gesture recognition capabilities. Surface electromyography (sEMG) signals, as an important type of bioelectric signal, can reflect human movement intentions. However, their non-stationary nature and low signal-to-noise ratio pose challenges for recognition algorithms. This study utilized the publicly available Ninapro DB1 dataset and employed a standardized preprocessing and feature extraction workflow to systematically compare the performance of four machine learning algorithms — Random Forest (RF), Multi-Layer Perceptron (MLP), LightGBM, and K-Nearest Neighbors (KNN) — in sEMG gesture recognition tasks. Experiments were conducted using 10-fold cross-validation, with performance evaluated across multiple metrics including accuracy, precision, recall, and F1 score. The results show that LightGBM performs best across all metrics (all exceeding 88.5%), demonstrating strong feature learning and generalization capabilities. This study provides empirical evidence for the selection of sEMG gesture recognition algorithms and offers guidance for the practical application of rehabilitation robot systems.
Weiliang Chen· Proceedings of the 3rd Inter...· 0 citations
Related blog posts
Microsoft Research Blog· microsoft.comAug 31, 2026
What if pathology foundation models could do more with less? GigaPath-Flash and GigaTIME-Flash cut computational demands while maintaining strong performance, opening the door to larger studies and broader exploration. The post GigaPath-Flash and GigaTIME-Flash: Toward population-scale discovery with efficient pathology foundation models appeared first on Microsoft Research.
MIT News · Artificial Intelligence· news.mit.eduAug 31, 2026
With millions of users across the world, Julia has been used to conduct cutting-edge research and to design new drugs, jet engines, heat pumps, and more.
MIT News · Artificial Intelligence· news.mit.eduAug 27, 2026
A new machine-learning framework aims to improve the success rate of computational protein design while moving away from results that reproduce sequences found in nature.