A Compact Neural Network for High Accuracy Bioimpedance-Based Hand Gesture Recognition

This paper presents a compact customised neural network with 44 parameters for hand gesture recognition based on electrical impedance tomography (EIT) using a flexible 8-electrode band. The classification accuracy is improved by assigning higher weights to the impedances captured closer to the current injection position. The non-fully connected layer working as a spatial filter reduces the complexity of the network structure. Validated on a discrete EIT system, the proposed network structure can distinguish eight gestures with an accuracy of 99.49%. Towards a low-power wearable design, an analogue inference circuit based on the proposed network structure was also designed in 65 nm CMOS. The system features a low-power multi-output digital-to-analogue converter (DAC) to provide data for the analogue computation efficiently. This designed CMOS analogue inference has a recognition accuracy of 98.13%.