QASHI: Quality-Aware Scale-Protected Hybrid Inference for Small-Object Detection in Optical and SAR Remote Sensing Imagery

Object detection in high-resolution remote sensing imagery faces a resolution–context trade-off: resizing weakens small objects, whereas cropping can truncate targets and remove context. We present QASHI, a detector-retraining-free hybrid inference framework combining a full-image view, four overlapping tiles, and at most three additional regions of interest (ROIs). Quality-Aware Density (QAD) prioritizes local re-inference by combining small-candidate density with weak responses, candidate competition, and full/tile prediction inconsistency. Graph-based ranking correction, bounded BoxVote refinement, and scale-protected routing integrate local predictions with full-image context. Experiments cover VisDrone2019-DET, DIOR, SeaDronesSee v2, and SRSDD-v1.0 with two RT-DETR detectors. On VisDrone-val with RT-DETRv4-S, QASHI improves AP by 2.84 points over the SAHI baseline while reducing mean latency by 14.40%. Relative to Full, it gains 5.74 AP points at 4.52 times the latency. Selector ablations and local parameter analyses characterize how region priority and fusion settings affect detection accuracy and computational cost. On SRSDD, display-domain image-quality sensitivity tests with RT-DETR-R50 yield higher mean perturbed AP (32.52 versus 31.41), with relative AP retention of 93.44% versus 93.98%. QASHI combines prediction-guided local processing and source-aware fusion to improve remote sensing detection through inference-time use of a fixed detector.