DHAF-YOLO: Dynamic Hierarchical Attention Fusion for Small Object Detection

Abstract

Small object detection in remote-sensing images remains challenging due to weak object features, scale variation, and complex background interference. To address these issues, this paper proposes DHAF-YOLO, a dynamic hierarchical attention fusion detector based on YOLOv11. The proposed method integrates three key designs: backbone enhancement, hierarchical cross-scale fusion, and adaptive focus refinement. First, Dynamic Weighted Feature Convolution (DWFC) is introduced to generate spatially sensitive responses and enhance fine-grained feature extraction. Second, Cross-Scale Gated Fusion (CSGF), implemented through dynamic scale-sequence fusion, performs two-stage interaction across P2–P5 pyramid levels to preserve high-resolution details and improve semantic consistency. Third, Dual-Path Adaptive Focus (DPAF) further refines fused features to strengthen target-related responses and suppress background noise. Experiments on the VEDAI-Coco dataset show that DHAF-YOLO achieves the highest Recall, mAP@0.5, and mAP@0.5:0.95 among compared YOLO-series detectors from YOLOv5 to YOLOv12, although its Precision is lower than YOLOv9. Compared with YOLOv9, DHAF-YOLO improves Recall by 2.4 percentage points and mAP@0.5:0.95 by 5.0 percentage points, while Precision decreases by 4.0 percentage points. These results demonstrate its effectiveness for remote-sensing small object detection.