Abstract:

Abstract—Convolutional neural networks (CNNs) are widely used in EMG-based motion recognition but typically incur high computational and memory costs, limiting deployment on edge or wearable platforms. To address this challenge, this paper proposes a lightweight CNN model based on mixed-precision quantization. By assigning different bit-widths (2/4/8 bits) to each network layer according to its sensitivity to precision, the proposed method significantly reduces both model size and inference complexity while maintaining classification performance. Specifically, we design a CNN model targeting five-class EMG gesture recognition based on the Ninapro DB1 dataset. The original model achieves an accuracy of 92.68% under FP32 precision, with a model size of 252.4 KB. By applying a layer-wise mixed-precision quantization strategy combining post-training quantization (PTQ) and quantization-aware training (QAT), the model is compressed to 36.58 KB (6.9× compression) with only a 3.1% accuracy drop. Hardware deployment on a RISC-V-based Milk-V Duo platform confirms the proposed approach’s suitability for embedded applications, achieving a 5.94× speedup over FP32 inference and a 1.65× improvement over INT8. These findings demonstrate the method’s potential for efficient, real-time EMG recognition in edge computing environments.