End-to-End Hand Gesture Recognition Based on Dynamic Graph Topology Generating Mechanism and Weighted Graph Isomorphism Network
Zidong Yu, Changhe Zhang, Xiaoyun Wang, Chao Deng · 2024
Hand gesture recognition (HGR) based on high-density surface electromyography (HD-sEMG) signals and the hybrid model of convolutional neural network (CNN) and graph neural network (GNN) has attracted extensive research interest in recent years. The key point of these methods is to correctly design the graph topology to establish the spatial relation between the information hidden in different electrodes (node features), and use the GNN to make up for the short-range dependence defect of CNN. Current methods determine the graph topology by prior knowledge or linear correlation of node features. As a result, the graph topology is static or unable to take into account nonlinear relation between node features, limiting the performance of the hybrid model. To address this problem, this study proposes a dynamic graph topology generating (DGTG) mechanism. DGTG adaptively and nonlinearly establishes graph topology by building nonlinear mapping between node features (extracted by CNN) and weighted adjacency matrix. At the same time, weighted graph isomorphism network (wGIN) is proposed to deal with graph data with edge weights. Finally, a hybrid network named DGTG-wGIN is formed. Experiments on a public dataset containing 65 hand gestures show that the proposed method achieves the state-of-the-art average accuracy of 96.63 ± 1.67 %. DGTG-wGIN innovatively integrates the determination of graph topology into the end-to-end learning process. Experiments show that this integration result in a graph topology that is closely related to hand gestures and sparse overall, contributing to the enhanced universality and generalization ability of the model.