An ATR Method for Imbalanced Data SAR Images Based on CCEGAN

Synthetic aperture radar (SAR) automatic target recognition (ATR) is an important application of SAR. In military applications, it is unrealistic to be able to acquire balanced SAR target images due to the non-cooperative nature of the targets. SAR ATR systems often face the problem of unbalanced data, which leads to the deterioration of the recognition performance of the ATR model. we propose a SAR ATR method based on a complementary cross-entropy generative adversarial network (CCEGAN), aiming to improve the robustness of the ATR model under imbalanced data. Firstly, features are extracted from the imbalanced data using a deep convolutional neural network (DCNN). Next, a classifiable GAN model is designed and then trained using features extracted from the DCNN. With the adversarial training of the GAN, the classification task is automatically accomplished while enriching the training sample set without the need for secondary screening of the generated samples. Finally, to further cope with the ATR model's preference for majority classes, we introduce CCE as the classification loss for classifiable GAN training. This will result in a relatively balanced weighting of minority and majority classes in the loss computation, reducing the dominance of the majority class over the overall loss. The method is an end-to-end model without manually designed features. Experiments on the measured MSTAR dataset validate the robustness of the method under different imbalance rates and different minority classes.