Attentional Convolutional Neural Networks Meet Kolmogorov-Arnold Networks for sEMG-Based Gesture Recognition

Surface electromyography (sEMG) reflects muscle contraction states, provides critical signals for decoding hand movements, and thus holds significant potential in applications such as human-computer interaction and prosthetic control. However, it still a significant challenge to model the complex, nonlinear, and dynamic patterns in the sEMG signals. In this paper, we propose a novel deep neural network that integrates convolutional neural network (CNN) and Kolmogorov-Arnold networks (KAN), termed ACNN-KAN, to utilize both CNN's local context modeling and KAN's adaptive nonlinearity for gesture recognition. Specifically, CNNs are first utilized to extract spatial features and an attention mechanism is integrated to enhance global dependency modeling. KAN is then leveraged to learn feature representation. We evaluate ACNN-KAN on three publicly available datasets (i.e., Ninapro DB1, DB5, and Myo) against nine competitors in terms of four performance metrics. Experimental results show that ACNN-KAN outperforms its competitors and achieves accuracies of 88.59% on DB1, 90.42% on DB5, and 96.10% on Myo, demonstrating its adaptability of KAN in gesture recognition.