Combining Local Electromagnetic Scattering and Global Structure Features for SAR Open Set Recognition

Qinjie Lin, Hao Sun, Yanjie Xu, Junyu Wang, Kefeng Ji, Gangyao Kuang

Abstract: Automatic target recognition (ATR) is crucial for synthetic aperture radar (SAR) image interpretation. However, existing SAR ATR primarily rely on algorithms from the field of computer vision, many methods have not adequately considered the imaging sensitivity of SAR images. Moreover, SAR ATR typically follows the closed set assumption, but in dynamic and open real-world environments, SAR systems may encounter target classes that have not been seen during training. To address these issues, this article proposes an SAR open set recognition (OSR) method that combines local electromagnetic scattering and global structure features. The global structural features focus on the overall relationship between target and background, while the local electromagnetic scattering features can effectively capture the scattering characteristics of target. Hence, we construct a global and local feature guidance network, fuses these complementary features as prior knowledge to guide deep features. Then, a class-anchor-based loss function is constructed by anchor loss and tuple loss, which cluster the features of known classes targets around predefined class anchors. Finally, the distance ratio is calculated to identify known classes targets and reject unknown classes targets. The experimental results on the MSTAR and SAR-ACD datasets show that our method outperforms other OSR methods in recall, precision, F1, and accuracy, demonstrating a superior performance in OSR.

External IDs:dblp:journals/staeors/LinSXWJK25