RAG-Net: Relational-Context Auxiliary Geometric-Regression Network for Tiny Object Detection in Remote Sensing Images

Abstract

Tiny object detection in remote sensing images is challenged by weak features, spatial detail loss, and extreme sensitivity to localization shifts. To tackle these issues, we propose RAG-Net, a unified detector built upon YOLOv12n. It leverages the Global Relational Context Hub module to reinforce weak features by integrating global relational aggregation with fine-grained local refinement, and a Decoupled DetectAux module to preserve mid-level feature details through auxiliary supervision. Additionally, the SIoU-based geometry-aware regression loss is adopted to mitigate localization drift by unifying angle, distance, and shape constraints. Experiments on four standard benchmarks validate the effectiveness of the three modules and their generalization across different detection frameworks. Furthermore, compared to related methods, RAG-Net achieves an optimal balance between accuracy and model complexity.