Improving small object detection in open water maritime monitoring under low-shot learning

Tools

Liao, Weiqiang and Ma, Qixiang and Wang, Haibin and Li, Huihui (2026) Improving small object detection in open water maritime monitoring under low-shot learning. Ocean Engineering, 343 (Part 3). 123307. ISSN 0029-8018 (https://doi.org/10.1016/j.oceaneng.2025.123307)

[thumbnail of Liao-etal-2025-Improving-small-object-detection-in-open-water-maritime-monitoring-under-low-shot-learning]
Preview
 Text. Filename: Liao-etal-2025-Improving-small-object-detection-in-open-water-maritime-monitoring-under-low-shot-learning.pdf
Accepted Author Manuscript
License: Creative Commons Attribution 4.0 logo
Download (7MB)| Preview

Abstract

Small object detection in open-water environments faces numerous challenges, including weak feature representation, limited annotated samples, imbalanced data distribution and complex background interference. To address these issues, this study proposes an enhanced YOLOv8-based detection algorithm for maritime monitoring scenarios. First, a full-network fine-tuning transfer learning strategy is employed to adapt pre-trained visual representations to low-shot conditions, thereby improving feature extraction efficiency. Second, a dynamic data augmentation approach based on Albumentations is introduced to expand the training sample space and enhance the model’s robustness to scale variations and deformation patterns. Finally, an optimized Focal and Efficient IOU (Focal-EIOU) loss function is incorporated into the bounding box regression process, integrating IOU, distance and angle losses while assigning higher weights to high-quality anchors, effectively balancing localization precision and sample imbalance. Experimental results on a real-world open-water dataset demonstrate that the proposed method improves mAP@0.5:0.95 by 8.3 % compared with the baseline model, with detection accuracy gains of 7.1 %, 8.3 % and 9.4 % for passerby, swimmer and suspected swimmer categories, respectively. To further validate generalization, additional experiments were conducted on the SeaDronesSee maritime dataset, where the proposed method outperformed the baseline across multiple categories. These results demonstrate that the proposed approach can substantially enhance small object detection under low-shot learning conditions and highlights its potential for practical applications in maritime rescue and waterway monitoring.

ORCID iDs

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