MERT-DETR: Multiorder Gated and Edge-Enhanced Transformer for Remote Sensing Small-Object Detection

Abstract

Renowned for its real-time detection capabilities, RT-DETR efficiently performs object detection in complex scenarios. However, small-object detection, particularly in remote sensing or maritime imagery, is frequently hindered by background interference, occlusion, and diminutive object features, thus limiting overall model performance. To address these issues, we propose MERT-DETR, an enhanced small-object detection framework built upon RT-DETR. The core of this framework relies on a newly designed multiorder spatial edge backbone network (MSE-Net), which progressively fuses edge information across multiple stages through its internal modules, substantially strengthening the fine-grained local details of small objects. Furthermore, to optimize cross-scale feature propagation and alleviate the localization challenges of tiny objects, the architecture introduces a multiscale attention interaction block (MSAI block) in the neck network. This is supplemented by a joint loss function (MIN Loss)—combining matchability-aware loss (MAL), Normalized Wasserstein Distance, and Inner-GIoU—to optimize the localization precision of tiny objects. Extensive experiments on the RSOD, NWPU VHR-10, and VisDrone2019 datasets demonstrate that MERT-DETR outperforms the baseline model, achieving mAP@0.5 scores of 96.2%, 92.8%, and 42.5%, respectively, while simultaneously reducing the model parameters and computational cost to 12.89M and 40.4 GFLOPs.