EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

DBPF-Net: A Dual-Branch Polarization Fusion Network for Floating Litter Detection from Unmanned Surface Vehicles

Jinting Zhu, Hui Cao, Cheng Fang

Journal of Marine Science and Engineering · 07.11.2026 · 14 (21) · 2111

DOI: 10.3390/jmse14212111 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Marine Science and Engineering, online erschienen am 07.11.2026. Verfasst von Jinting Zhu und 2 weiteren Autor:innen. Die Arbeit nennt 44 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Department of Transport of Zhejiang Province, Zhejiang Province’s “14th Five-Year Plan” Teaching Reform Projects for Higher Vocational Education.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Accurate detection of floating litter is important for autonomous monitoring and cleanup by unmanned surface vehicles (USVs), yet water-surface reflections, illumination variations, and small distant targets challenge RGB perception. Polarization provides complementary information, motivating fusion strategies that exploit the characteristics of both modalities. This study proposes the Dual-Branch Polarization Fusion Network (DBPF-Net) for floating litter detection. A Multiscale Appearance Block aggregates appearance cues in the RGB branch, while an Adaptive Frequency Block spatially modulates frequency components in the polarization branch. A High-Resolution Fusion Branch combines shallow RGB and polarization features through gated joint modeling at P2 to preserve fine spatial information. A Multiscale Context Block integrates contextual features at P5, and a bidirectional neck supports four-scale detection. Evaluated using the PoTATO plastic-bottle image dataset with the COCO evaluation protocol, DBPF-Net achieves 50.1% average precision (AP), 58.5% average recall (AR), and 46.0% AP for small objects, exceeding the dual-branch RGB–polarization baseline by 5.8, 6.3, and 8.2 percentage points, respectively. The model runs at 67.1 FPS on an NVIDIA RTX 4070 GPU. These results support the effectiveness of modality-specific enhancement and scale-aware fusion for overall and small-object detection in the evaluated water-surface scenes, while maintaining real-time inference on the experimental platform.

Förderung

  • Department of Transport of Zhejiang Province
  • Zhejiang Province’s “14th Five-Year Plan” Teaching Reform Projects for Higher Vocational Education

Zitieren

APA

Zhu, J., Cao, H., & Fang, C. (2026). DBPF-Net: A Dual-Branch Polarization Fusion Network for Floating Litter Detection from Unmanned Surface Vehicles. Journal of Marine Science and Engineering, 14(21), 2111. https://doi.org/10.3390/jmse14212111

Vancouver

Zhu J, Cao H, Fang C. DBPF-Net: A Dual-Branch Polarization Fusion Network for Floating Litter Detection from Unmanned Surface Vehicles. Journal of Marine Science and Engineering. 2026;14(21):2111. doi:10.3390/jmse14212111

BibTeX
@article{Zhu2026,
  author = {Zhu, Jinting and Cao, Hui and Fang, Cheng},
  title = {DBPF-Net: A Dual-Branch Polarization Fusion Network for Floating Litter Detection from Unmanned Surface Vehicles},
  journal = {Journal of Marine Science and Engineering},
  year = {2026},
  volume = {14},
  number = {21},
  pages = {2111},
  doi = {10.3390/jmse14212111}
}
Kostenlos starten →