All-ConvNet: A Lightweight All CNN for Neuromuscular Activity Recognition Using Instantaneous High-Density Surface EMG Images

Abstract

Neuromuscular activity recognition using low- resolution instantaneous high-density surface electromyography (HD-sEMG) images present a great challenge. The recent result shows the high potentiality and hence opens up new avenues for the development of more fluid and natural muscle-computer interfaces. However, the existing approaches employed a very large deep ConvNet, which requires learning >5.63 million training parameters only during fine-tuning and pre-trained on a very large-scale labeled HD-sEMG training datasets, as a result, it makes high-end resource bounded and computationally expensive. To overcome this problem, we propose a lightweight All-ConvNet model that consists solely of convolutional layers, a simple yet efficient framework for learning instantaneous HD-sEMG images from scratch through random initialization. Without using any pre-trained models, our proposed lightweight All-ConvNet demonstrate very competitive or even state of the art performance on a current benchmarks HD-sEMG dataset, while requires learning only ~460k training parameters and using ~12xsmaller dataset. The experimental results proved that the proposed lightweight All-ConvNet is highly effective for learning discriminative features for low-resolution instantaneous HD-sEMG image recognition and low-latency processing especially in the data and high-end resource constrained scenarios.