"20260416_K-C-Lee+Y-H-Hsu"

"Limited availability of labeled training data remains a critical challenge for learning based recognition systems, often resulting in reduced accuracy and poor generalization. In particular, the SAR (synthetic aperture radar) image is often difficult to obtain.To address this issue, this paper explores generative data augmentation as an effective solution under limited data conditions. A hybrid generative model termed WACGAN-GP is constructed by integrating auxiliary class supervision with the Wasserstein adversarial framework and gradient penalty. This design enables stable training and ensures class consistent sample generation. The synthesized samples are incorporated into the training process to enhance data diversity and improve classifier performance. Experimental evaluations are conducted on the MSTAR dataset and the OpenSARShip dataset under limited sample settings, considering both balanced and imbalanced data distributions. The proposed augmentation strategy is compared with conventional transformation based methods, including horizontal and vertical flipping. The results show that data augmentation using WACGAN-GP consistently improves recognition accuracy and outperforms conventional augmentation methods. These findings confirm its effectiveness under limited training data conditions."