EMe-DETR: an efficient and multi-scale enhanced solution for multi-object detection in optical remote sensing images

Object detection in optical remote sensing images, particularly leveraging vision transformers (ViTs), has achieved significant progress. However, the small object detection remains a primary challenge. The small size of objects in optical remote sensing images impedes the extraction of fine-grained features, leading to difficulties in accurate identification and classification. Moreover, the significant overlap among dense objects exacerbates difficulties in precise localization and categorization. Concurrently, conventional detection networks often struggle to balance recognition accuracy with model complexity. To address these issues, we propose the EMe-DETR, an efficient and multi-scale enhanced network based on a modified real-time detection transformer (RT-DETR). First, we introduce an efficient feature-aware interaction module (EFIM) that significantly enhances the capability to extract critical features from complex backgrounds. Furthermore, a multi-scale enhancement pyramid is proposed with a lightweight design to reduce complexity while preserving high performance. Experimental results demonstrate that the EMe-DETR achieves mean average precision (mAP) scores of 88.43% and 94.90% on the DIOR and HRRSD datasets, respectively. Additional validation confirms that the EMe-DETR effectively balances accuracy and efficiency, offering a robust and innovative solution for object detection in optical remote sensing images.