RIF-YOLO-N: Lightweight P3 Residual Identity Fusion with Resolution-Guided Fine-Tuning for Tiny-Object Detection

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Tiny-object detection remains challenging because repeated downsampling degrades the limited spatial information available for small targets. This paper proposes RIF-YOLO-N, a lightweight detector that strengthens the finest P3 pathway of YOLOv8n through residual identity fusion (RIF). RIF combines neck-level semantic features with backbone-level spatial features via an identity-safe residual mechanism, preserving the original anchor-free P3/P4/P5 detection hierarchy. Resolution-guided fine-tuning (RGFT) further adapts the detector to higher-resolution inputs without adding a new detection head. The framework is evaluated on VisDrone, UAVDT, and VEDAI. At 1024×1024, RIF-YOLO-N with RGFT improves mAP50:95 from 0.235 to 0.243 on VisDrone, from 0.519 to 0.543 on UAVDT, and from 0.252 to 0.419 on VEDAI. On VisDrone, mAP50 also increases from 0.401 to 0.415, while model complexity rises only from 3.008M to 3.061M parameters and from 8.1 to 8.8 GFLOPs. Size-stratified analysis shows improvements for both tiny and small objects, with the larger AP50 gain observed for the small-object group. Ablation and resolution studies further support the selected P3 enhancement and show that RGFT performance depends on the chosen fine-tuning resolution. Overall, RIF-YOLO-N improves lightweight tiny-object detection without introducing an additional prediction scale or substantially redesigning the detector.