DGL-YOLO: detail-enhanced and gated-fusion YOLO11 for efficient remote sensing object detection

Abstract

Object detection in remote sensing imagery (RSI) underpins UAV monitoring, precision agriculture, and disaster assessment, yet is hindered by extreme scale variation, vast homogeneous backgrounds, and small target footprints. Existing lightweight YOLO-based detectors typically improve the backbone attention, downsampling, or neck fusion in isolation rather than jointly addressing channel/spatial redundancy, downsampling detail loss, and coarse cross-level fusion. This paper proposes three complementary YOLO11-based modules: (1) C3k2-LWGA, which partitions channels into point-wise, local, medium-range, and global pathways to capture RSI scale diversity; (2) DEAG, which combines pinwheel convolution, C3k2-LWGA refinement, and context anchor attention via dual multiplicative gating to preserve fine detail and inject long-range context during downsampling; and (3) DPCF, which replaces plain neck concatenation with a learnable gated fusion of low- and high-level features. Evaluated on VEDAI and RSOD, a full factorial ablation shows that each module contributes independently and the full model improves mAP50/mAP50:95 from 60.4%/35.4% to 63.3%/37.5% at a modest cost increase (8.1→10.8 GFLOPs; 3.007M→3.400M parameters). Compared with four lightweight RSI detectors, the proposed model achieves the smallest parameter count (3.4M), the best accuracy on VEDAI, and the second-best accuracy on RSOD, demonstrating an improved accuracy–parameter trade-off for small, densely-distributed targets.