Single Gate Spiking Recurrent Neural Network for Histogram-less Single-Photon Depth Sensing

Recent research has shown the potential of directly coupling time-resolved single-photon avalanche diode (SPAD) sensors with neural networks to streamline signal processing and reduce redundancy. However, prior attempts have been limited to general neural network models, including Long Short-Term Memory (LSTM), which fall short of task-specific optimization. Based on insights into the demands of the SPAD depth sensing task, this work introduces a highly simplified spiking recurrent neural network tailored for this purpose. This model features a unique single input gate architecture implemented solely with simple spiking neurons. It demonstrates improved accuracy over classic LSTM with a 1.95x reduction in parameter count and at least an 8.40x reduction in energy consumption while significantly outperforming other spiking neural networks. The results underscore the significance of developing network architectures dedicated to current tasks, which could pave the way for potential advancements in fully in-pixel processing.