FX-DETR: A Lightweight and Efficient Model for Maritime Small Object Detection

This paper proposes a novel end-to-end object detection algorithm named FX-DETR, aiming to enhance detection efficiency while reducing computational complexity. Based on RT-DETR, the method introduces the lightweight FasterNet as the backbone network, effectively decreasing the number of parameters and FLOPs. Moreover, to address the upsampling process in the Cross-Scale Feature Fusion Module (CCFM) of RT-DETR, this study replaces it with the content-aware up-sampling operator CARAFE, thereby further improving feature representation capabilities. While maintaining detection accuracy, the proposed approach significantly reduces computational overhead. Experiments conducted on the SeadroneSee dataset demonstrate that FX-DETR achieves 80.5%mAP@0.5with only 28.8 GFLOPs, striking a better balance between efficiency and accuracy compared to the baseline model (58.3 GFLOPs, 81.9%mAP@0.5).This represents a 50.6% reduction in computational cost with only a 1.4% drop in accuracy, highlighting the effectiveness and deployment potential of the proposed improvements. These improvements confirm the potential of the proposed method for maritime small object detection applications.