MOSR: An Open-Set Recognition Network Based on Masked Autoencoder for Ship Detection

In remote sensing image classification, open-set recognition (OSR) poses a significant challenge, aiming to accurately classify known categories while effectively rejecting unknown class samples or identifying potential novel categories. Although existing methods have made strides in recognizing known classes, they exhibit notable limitations in handling unknown class samples. This letter introduces an OSR model for ship detection, termed masked autoencoder (MAE)-based OSR (MOSR), which leverages the robust representation learning capabilities of the MAE. MOSR not only sustains high accuracy in the recognition of known classes but also markedly enhances the performance in the identification of unknown class samples. Comprehensive experiments on the custom RSHIP-137 remote sensing dataset validate the efficacy and superiority of the MOSR model. Compared with the state-of-the-art (SOTA) adversarial reciprocal point learning (ARPL) method, MOSR shows substantial improvements in both known class recognition accuracy and the area under the receiver operating characteristic curve (AUROC) for unknown class recognition for ship detection. This study presents a novel solution for OSR in remote sensing ship detection and offers valuable insights for future research.