Small-Data Gesture Recognition Using Bending-Sensitive Graphene Strain Sensors and Diffractive Deep Neural Networks
瑾 李 · Modeling and Simulation · 2025
人工智能识别手势在认知神经科学和机械臂技术领域有着广泛的应用。最近，摄像头和应变传感器被用于收集不同手势的图片和随时间变化的响应数据，这些数据被输入人工神经网络进行识别。大量的输入数据会造成算力的浪费、大量的能源消耗和明显的时间延迟。文章介绍了一种基于弯曲敏感石墨烯应变传感器和衍射深度神经网络(Diffraction Deep Neural Network, D2NN)的手势识别方法。应变传感器是通过激光划线技术在氧化石墨烯(Graphene Oxide, GO)薄膜上制备的，在弯曲角度调节下表现出可调的电流响应。10个应变传感器位于手指关节处，用于捕捉不同手势下的电流响应，并将其转换为10像素图像，作为衍射深度神经网络的输入。在3次迭代内，识别0~9数字手势的仿真准确率可达到100%。输入数据量仅为10个，远低于已报道的手势识别设备。石墨烯应变传感器与光学神经形态计算的结合，为实现低成本、高效率、高准确度的人机交互铺平了道路。Artificial intelligence recognition of gestures shows a wide range of applications in cognitive neuroscience and robotic arm technology. Cameras and strain sensors have recently been used to collect photo and time-dependent response data for different gestures. The data are input into artificial neural networks for recognition. The large amount of the input data induces computing power waste, massive energy consumption, and significant time delay. This paper introduces a gesture recognition pathway based on bend-sensitive graphene strain sensors and diffractive deep neural network (D2NN). The strain sensor is fabricated on graphene oxide (GO) film through laser scribing and exhibits tunable current responses under the adjustment of bending angles. 10 strain sensors are located at finger joints to capture the current responses under different gestures and convert them into a 10-pixel image as the input of diffractive deep neural networks. 100% accuracy can be achieved within 3 iterations to recognize gestures representing numbers 0~9. The amount of input data is only ten, much less than those of the reported gesture recognition devices. The pathway combining graphene strain sensors with optical neuromorphic computing paves the way for achieving low-cost, efficient, and high-accuracy human-computer interaction.