Classifying LPI signals with transfer learning on CNN architectures

Due to the increased deployment of low probability of intercept radar systems, recognition and classification of low probability of intercept signals has developed an increased importance for electronic warfare systems. Recent results showed that combining time-frequency transformations such as Choi-Williams distribution with convolutional neural networks yield high accuracy. Since training convolutional neural networks is a time consuming task, we propose to use transfer learning on pre-trained convolutional neural network architectures. Furthermore, we compare these retrained neural network to the neural network trained with randomly initialized weights. We will demonstrate that the required training time for the transfer learning method is significantly shorter. Moreover, classifying time-frequency images based on Choi-Williams distribution achieves for both weight initialization methods an accuracy of over 99%.