ArgusSAR: A Multidimensional Perception and Robust Detection Method for Multiscale Objects in SAR Images

Due to the challenges presented by the different scales of objects and the complex interference in synthetic aperture radar (SAR) images, current detectors struggle to achieve satisfactory accuracy and exhibit weak generalization abilities across different datasets. This paper proposes a robust and universal object detector framework named ArgusSAR, which is suitable for both ship and aircraft object detection. The proposed framework comprises three key components: In detail, a Spatial Encoder, which employs Multi-Head Attention Blocks integrated with multi-layered Encoder Blocks, extracts robust feature representations while suppressing irrelevant interference inherent in complex SAR backgrounds. Moreover, a Fine-grained Feature Aggregation Encoder, which incorporates the multi-level pyramid architecture with content-aware upsampling operations, enabling precise integration of multi-dimensional features and enhanced extraction of object boundaries. Furthermore, a robust Multi-dimensional Awareness Attention Decoder, which incorporates scale-spatial-task self-attention mechanisms to strengthen geometric perception and improve both localization accuracy and generalization ability for ship and aircraft objects. Three publicly available SAR datasets involving ship and aircraft objects are chosen to confirm the effectiveness of ArgusSAR. Compared with the baseline and the other methods, our method has the best performance on these datasets. It demonstrates that ArgusSAR framework is robust for SAR ship and aircraft objects detection. The source code is available at https://github.com/masurq/ArgusSAR.