SSC-DETR: Real-Time End-to-End Small-Object Detection in UAV Remote Sensing Imagery via Saliency-Guided Sparse Attention and Signed Sparse Semantic Calibration

Small-object detection in UAV imagery is difficult because targets occupy only a few pixels amid cluttered backgrounds. Although high-resolution features preserve fragile spatial cues, dense contextual interaction over these features is computationally expensive, while multiscale fusion can weaken or misalign local semantic responses. We propose SSC-DETR, a real-time end-to-end detector that addresses these challenges through selective contextual interaction before fusion and selective semantic calibration afterward. Saliency-Guided Sparse Attention (SGSA) learns a spatial selection prior—a saliency map explicitly supervised by Gaussian targets derived from ground-truth bounding boxes—and uses it to select object-relevant high-resolution tokens for sparse self-attention while retaining dense local features elsewhere. Signed Sparse Semantic Calibration (SSSC) independently selects a smaller set of post-fusion positions and injects spatially aligned, bounded signed semantic residuals to reinforce useful evidence or suppress inconsistent responses. Haar-DWT downsampling and a scale-calibrated GIoU–GCD localization objective complement the two modules. Experiments on VisDrone and AI-TOD-v2 demonstrate improvements with ResNet-18 and ResNet-50. With ResNet-50, SSC-DETR increases AP from 29.6% to 35.2% on VisDrone and from 25.5% to 28.9% on AI-TOD-v2 while retaining real-time inference.