Ship Recognition Method Based on Spiking Neural Networks in Synthetic Aperture Radar Images

Abstract

Ship recognition is a hot issue for the intelligent interpretation in the synthetic aperture radar (SAR) images. A modified ship recognition approach based on the spiking neural networks (SNNs) in the SAR images has been proposed. Firstly, a Poisson encoder on the basis of the visual attention mechanism is used to encode the input SAR image in a pulse sequence, to remove the background noise interference during the encoding process and retain the visual saliency of the image as much as possible. Then, based on the neuron model of the leaky integrate and fire (LIF), combined with the convolutional neural network, an end-to-end SNNs model was constructed, and the LIF neuron pulse emission frequency has been used to perform the target recognition in the SAR images. Finally, to resolve the issue that the SNNs model is very difficult to train, an arctangent function is used as a substitute function of the pulse emission function during backpropagation for the gradient calculation, and the backpropagation algorithm is applied to the training process of the ship target recognition. The experimental results prove that the presented approach can accurately identify the ship targets and has the advantages of fewer parameters and high efficiency.