Multi Scale SAR Image Target Detection Algorithm Based on RT-DETR

To address low detection accuracy in SAR images due to complex backgrounds, large scale variations, and small targets, an improved algorithm SRA-DETR based on RT-DETR is proposed. It uses repBlock on four backbone branches to reduce parameters and enhance speed, incorporates a self-attention vision transformer for feature similarity extraction, and employs SE-CBAM for small target detection to suppress background interference. Adaptive path downsampling improves feature retention and background filtering compared to nearest neighbor methods. SRA-DETR decreases parameters by 40% and computational complexity by 20%, achieving aP50 of 96.6% and increasing mAP from 77.9% to 80.1% over the original RT-DETR.