Conversion of CNN to SNN for Aerospace Object Detection Tasks

Abstract

Spiking Neural Networks (SNNs) is regarded as the next generation of Artificial Neural Networks (ANNs) and an alternative to Convolutional Neural Networks (CNNs). The CNN– SNN conversion is considered one of the most promising approaches to generate SNNs. However, previous works showed that the conversion methods required a long timesteps (also known as inference latency in CNNs) to get an acceptable accuracy loss. On the other hand, SNN is rarely used for object detection. In this paper we proposed a Leaky-Integrate-and-Fire (LIF) neuron model which can complete convolution and FC operations, and based on the neuron model we convert the widely used Yolov3-tiny into SNN, then validate our method by detecting crater of Martian and Lunar, which are of vital significance in various space exploration missions such as spacecraft auto-landing task in deep space. The experiment results shows that our conversion method can detect the object within short timesteps, and the objects were first detected at a time step of 24 and more objects were detected as the timestep increased. After expanding the timesteps to 128, it reported precision 0.504, recall 0.453, mAP@0.5 0.396, and F1 0.477. To the best of our knowledge, this is the first time to use SNN in the field of space object detection.