A Lightweight Real-Time Remote Sensing Small Object Detection Network Based on Dynamic Group Aggregation and Multiscale Feature Interaction

Real-time small object detection in remote sensing is frequently hindered by high-density distributions and occlusions, which complicate effective multiscale feature fusion. To reduce false positives and false negatives, we propose S-RT-DETR, a real-time detection framework tailored for small objects in remote sensing images. A dynamic complementary reparameterization block (DCRB) is introduced into the ResNet-18 backbone to improve feature extraction through dynamic group-shuffling and multiscale convolutional aggregation. An adaptive multiscale intrafeature attention (MIFA) module incorporates adaptive region attention (ARA) and hierarchical fusion to strengthen interactions among spatial, channel, and semantic features. Furthermore, an enhanced content-guided attention fusion (CGAF-E) module enables lightweight multiscale representation and improves discrimination of visually similar objects. Experiments on SIMD show that S-RT-DETR achieves an mAP50 of 81.2%, outperforming the baseline by 4.4%, 5.3%, and 4.2% in precision, recall, and mAP50, respectively, while reducing parameters to 15 M and computational cost to 49 giga floating-point operations (GFLOPs). On VisDrone2019 and HRSC2016, S-RT-DETR consistently demonstrates superior detection accuracy, efficiency, and lightweightness. Furthermore, heatmap and receptive-field analyses validate the model's enhanced focus on small objects and robustness under complex scenarios.