SODFPN-YOLO : a feature pyramid-enhanced method for detecting small objects on ocean surface

Tools

Ma, Qixiang and Liao, Weiqiang and Wang, Haibin and Dong, Xin and Li, Huihui (2026) SODFPN-YOLO : a feature pyramid-enhanced method for detecting small objects on ocean surface. Ocean Engineering, 357 (Part 2). 125506. ISSN 0029-8018 (https://doi.org/10.1016/j.oceaneng.2026.125506)

[thumbnail of Ma-etal-OE-2026-SODFPN-YOLO-a-feature-pyramid-enhanced-method-for-detecting]
Preview
 Text. Filename: Ma-etal-OE-2026-SODFPN-YOLO-a-feature-pyramid-enhanced-method-for-detecting.pdf
Accepted Author Manuscript
License: Creative Commons Attribution 4.0 logo
Download (41MB)| Preview

Abstract

Ocean surface object detection is prone to clutter interference, a challenge that is even more pronounced for small objects due to their small size, weak feature representation, and blurred boundaries. The lack of high-quality datasets exacerbates these difficulties under complex maritime conditions. To address these issues, this paper proposes a YOLOv8-based feature enhancement framework, called SODFPN (Small Object Detection Feature Pyramid Network). Building on the traditional PAFPN structure, we introduce the SPDConv (Spatial Distribution Convolution) module to compress and semantically enhance P2 high-resolution features before fusing them with the P3 layer, thereby improving small object representation. In neck network, we integrate the OmniKernel module, which incorporates local, large receptive field, and global branches to capture both global context and local details at multiple scales, effectively mitigating semantic weakening and boundary blurring of small objects. Additionally, a CSP-OmniKernel module is proposed by combining the Cross Stage Partial (CSP) structure with OmniKernel, enabling efficient deep feature modeling and redundancy suppression. A dedicated maritime small object detection dataset, SeaTinySet, consisting of 2420 precisely annotated images, is also developed. Experimental results demonstrate that the proposed method outperforms several YOLO variants, achieving improvements of 1.2%–4.4% in mAP50 and 2.5%–15.8% in mAP50-95, confirming its effectiveness for accurate small object detection in complex ocean surface conditions.

ORCID iDs

Ma, Qixiang, Liao, Weiqiang, Wang, Haibin ORCID logoORCID: https://orcid.org/0000-0002-3520-6856, Dong, Xin and Li, Huihui;
CORE (COnnecting REpositories)