Semantic Assistance in SAR Object Detection: A Mask-Guided Approach

The unique challenge in SAR object detection is the strong speckle noise inherent in SAR imagery. Existing learning-based works mainly focus on architectural enhancements, and fail to consider the valuable semantic information that can mitigate the effects of speckle noise. Large pretrained segment anything model (SAM) is a powerful foundational model with general semantic knowledge. However, SAM is not fully exploited for SAR object detection. This study paves the way for applying SAM for SAR object detection. Rather than fine-tuning the SAM network, we propose three mask-guided learning strategies by simply utilizing the semantic masks generated by SAM. Built upon the advanced RealTime DEtection TRansformer (RT-DETR) framework, the Semantic Assisted DETR, deemed as SA-DETR, integrates prior semantics from SAM into the SAR detection task. To be specific, first, we propose the mask-guided feature denoising module in the encoder stage, to enhance the network's discrimination of positives and negatives. Second, we propose the mask-guided query selection for initial query generation, which is beneficial for the decoder refinement. Finally, the mask-guided instance segmentation is proposed to achieve more accurate localization. To validate the superiority of the proposed SA-DETR, extensive experiments are conducted on two benchmark datasets, i.e., the SAR ship detection dataset (SSDD) and the recently published COCO-level large-scale multiclass SAR object detection dataset (SARDet-100K). Experimental results on both datasets outperform previous advanced detectors, achieving a new state-of-the-art with 99.0 $AP_{50}$ and 88.4 $mAP_{50}$ on SSDD and SARDet-100 K, respectively.