Abstract: The detection of rocky obstacles onboard in the deep space environment is an important prerequisite to ensure the safe detection of the planetary rover.Due to the storage capacity and data processing capabilities of space-borne computing equipment,large-scale and complex calculations are not suitable for the remote and deep space environment.In addition,traditional rock detection algorithms have problems such as high complexity and excessive energy consumption.Therefore,this paper proposes the Spiking-Unet,which is a multi-class semantic segmentation algorithm and uses deep spiking neural network to achieve effective detection of rocks onboard.Firstly,because of class imbalance in the rock images,constructing the lovasz_CE loss function to train the Unet network model.Secondly,mapping the parameters obtaining from the Unet network model to the Spiking-Unet network based on the parameter scaling method.Thirdly,using the S-softmax function based on the pulse firing frequency to rea-lize the pixel-level classification of rock images.The proposed algorithm is tested on the public datasets Artificial Lunar Landscape.Experimental results show that the Spiking-Unet can reduce Flopsto about 1/1 000 of the original and reduce energy consuptionto about 1/600 of the original when the accuracy is similar with the Unet model with the same topology.

Key words: Deep space exploration, Spiking neural network, Rock detection, Image segmentation, Semantic segmentation