Ground Surveillance Radar Target Classification Based on 2D Selective CA CNN

In this paper, a new 2D convolutional neural networks (CNN) model is proposed for classifying targets detected by low-resolution ground surveillance radar. The targets echoes can be disposed in different domains to obtain their unique features which can be used in classification. Using the proposed network for radar target classification consists of two steps, using Toeplitz matrix to reconstruct original 1D signal to 2D and constructing the Selective CA (coordinate attention) model for target classification. In first step, Toeplitz matrix is made use of reconstructing Radar signal to build 2D data without changing the characteristic distribution of the signal, which makes it possible to train radar signal with 2D CNN. In second step, we construct the 2D Selective CA model. Each block of the model we propose applies the CA to different receptive fields for feature fusion. The residual structure is also included, which can make the information spread more smoothly in the network and effectively solve the problem of vanishing/exploding gradients int the back propagation process. The model was trained and tested on actual collected database containing human and car, which finally achieve 99.2% accuracy on the original frequency domain test set and 97.89% accuracy on the original time domain test set.