Edge Information Extraction Based Open-Set Airplane Detection in Remote Sensing Images

This article addresses the challenge of open-set airplane detection in remote sensing images, where the model must identify both trained known and untrained unknown target classes in dynamic environments. Considering the complex background and the low resolution of the targets makes it difficult to generate high-quality pseudolabels for the corresponding locations, we propose an edge information extraction-based open-set target detection (EI-OSTD) framework that enhances the detection of unknown classes by incorporating edge features into the detection process. The EI-OSTD framework includes two key components as follows. 1) An adaptive preselection module that optimizes candidate boxes for known classes using encoder output features, improving detection accuracy. 2) A pseudolabel selection strategy that leverages edge information to generate high-quality pseudolabels for unknown classes, thereby improving the recall of unseen targets. Experiments on the MAR20 and SAR-AIRcraft-1.0 datasets demonstrate that EI-OSTD not only maintains strong performance in detecting known classes but also significantly outperforms existing methods in identifying unknown classes.