RNN with stacked architecture for SEMG based sequence-to-sequence hand gesture recognition

Having proven their suitability for real world applications, surface electromyography signals are the means of choice for hand gesture recognition especially in medical applications like upper limb prosthesis. So far, mostly hand-crafted features combined with a standard classifier or neural networks are adopted for signal analysis. However, the performance of the standard approaches is insufficient and the networks are inappropriate for embedded applications due to their sheer size. To address these problems, a small recurrent neural network is proposed to fully utilize the sequential nature of the biosignals. Our network architecture features a special recurrent neural network cell for feature learning and extraction instead of convolutional layers and another type of cell for further processing. To evaluate the suitability of this inhomogenously stacked recurrent neural network, experiments on three different databases were conducted. The results reveal that this small network significantly outperforms state-of-the-art systems and sets new records. In addition, we demonstrate that it is possible to achieve relatively equal performance across all subjects.