Abstract
<title>Abstract</title> <p>Transformer-based detection methods offer global modeling capabilities for SAR imagery, yet they struggle with speckle noise, complex backgrounds, and high feature similarity in dense small target scenarios, often resulting in reduced accuracy and increased false alarms. To overcome these limitations, we introduce FMA-DETR, an adaptive Transformer framework specifically tailored for dense small target detection in SAR images. We propose a Feature Refinement Network (FRNet) as the backbone for fine-grained feature extraction in complex scenes. Between the backbone and neck, we incorporate a Multi-scale Spatial-Channel Mixed Attention (MSCA) module that refines backbone features via multi-scale interactions and channel attention, boosting discrimination of densely distributed small targets. In the neck, we develop an Adaptive Frequency-aware Fusion (AdaFreq) module that applies frequency-domain decomposition and adaptive resampling for efficient multi-scale fusion, mitigating the attenuation of small-target features. Experimental results demonstrate FMA-DETR's superior performance, achieving mAP\((_{50})\) of 90.7% on MSAR-1.0, 97.8% on SSDD, and 93.5% on HRSID.</p>