ADCG: A Cross-Modality Domain Transfer Learning Method for Synthetic Aperture Radar in Ship Automatic Target Recognition

Due to the powerful feature extraction and expression ability of convolutional neural networks (CNNs), exceptional success has been achieved in the field of ship automatic target recognition (ATR) of synthetic aperture radar (SAR). However, the CNNs cannot work effectively with sparse labeled samples and imbalanced categories. This study proposes a new attention-dense-CycleGAN (ADCG) method that is suitable for the ship transfer learning task from optical to SAR (OPT2SAR). The key improvement of the ADCG lies in the construction of a dense connection module (DCM) and a lightweight convolutional block attention module (CBAM). The DCM is able to overcome the problems of generator feature redundancy, large network model parameters, and severe training time in the original CycleGAN network. The lightweight CBAM can solve the problem of not being able to locate the main features of ships with a minimal increase in network parameters. Compared with the performance of other popular generative adversarial networks, the superior performance of the ADCG in the OPT2SAR transfer learning is demonstrated with the Frechet inception distance (FID) minimum of 76.04 and the Kernel inception distance (KID) minimum of 0.0403. Finally, the ability of pseudo-SAR domain images was tested to improve the recognition accuracy of popular ship classification networks, and this achieved an average improvement of 6% in recognition accuracy. Therefore, the results of this study verify the rationality, validity, and application value of pseudo-SAR domain in solving the problems of sparse marker samples and class imbalance in ship ATR network model.