Sar Image Classification Based on Spiking Neural Network with Different Coding Methods

Deep learning methods are widely used in synthetic aperture radar (SAR) image recognition and have achieved high recognition accuracy on some public datasets. However, the existing deep learning methods do not really form a cognitive level of understanding and are highly susceptible to spoofing attacks by external noise. Spiking neural network (SNN) contains the inherent spatiotemporal dynamics of biological neurons and has strong robustness. SAR image recognition task is accomplished based on SNN model using different coding methods in this paper, and the effects of different coding methods on SNN model are analysed in detail through different degrees of interference samples. Firstly, Poisson coding and direct coding methods are adopted to achieve spiking coding of SAR images. Then different SNN models are constructed for SAR image recognition based on different coding methods. Surrogate gradient method is then employed to solve the problem of non-differentiable spiking sequences, which effectively achieves the end-to-end training of SNN. Finally, efficient SAR image recognition is achieved based on the constructed SNN models. And the salt-and-pepper (S&P) noise samples and fast gradient sign method (FGSM) adversarial samples are produced to test the SNN models, which are used to compare and analyse the effects of different coding methods on the model performance. In this paper, the model is validated using the public dataset MSTAR. The experimental results show that Poisson coding is more likely to improve the robustness of the SNN model, and direct coding is more conducive to improve the recognition accuracy of the SNN model, which provides valuable insights for the design of SNN models in the future.