DGF-YOLO: a multi-scale feature fusion and refinement model for small floating debris detection in complex water-surface scenes

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Accurate detection of small floating debris on water surfaces is important for automated inspection and aquatic environment analysis. However, small floating debris often exhibits weak visual cues and low contrast, and its responses are easily disturbed by ripples, reflections, and shoreline clutter in complex water-surface scenes. To address these challenges, this paper proposes DGF-YOLO, an enhanced YOLOv12n-based framework that coordinates three complementary components to strengthen the high-resolution feature used for small-object prediction. The Guided Complementary Fusion Module (GCF) replaces the A2C2f blocks at selected stages of the YOLOv12n backbone and neck to enhance weak local object cues. The Scale-Sequence Fusion Module (SSFM) models ordered interactions among multi-scale backbone features and produces a high-resolution contextual representation. The Dual-branch Gated Refinement Module (DGRM) fuses this contextual representation with the neck-fused high-resolution feature before the detection head, thereby refining the prediction feature for small objects. Experiments on the FloW-Img and IWHR_AI_Lable_Floater_V1 datasets show that DGF-YOLO improves the mean mAP50 and mAP50-95 over the YOLOv12n baseline by 3.3% and 2.0% on FloW-Img, and by 3.0% and 3.7% on IWHR_AI_Lable_Floater_V1, respectively. Ablation and visualization analyses further show that the proposed design improves detection completeness, localization reliability, and object-related responses under complex water-surface backgrounds.