Bioinspired in-sensor vision processing network for action recognition

Recognizing actions from dynamic images is vital for numerous real-world applications, encompassing areas such as surveillance, autonomous vehicles, and human-computer interaction. However, current machine vision systems rely on complex circuits and algorithms. Here, we demonstrate an insect-inspired compact neural network with only one synaptic layer for in-sensor recognizing actions. MoS <inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> phototransistor can perceive and nonlinearly manipulate temporal information. With 20×20 phototransistor arrays, we have realized in-sensor compressing dynamic vision with >99% recognizing accuracy.