Open Access

ARTICLE

OD-YOLOv8: A Lightweight and Enhanced New Algorithm for Ship Detection

Zhuowei Wang^1,*, Dezhi Han¹, Bing Han², Zhongdai Wu²

1 School of Information Engineering, Shanghai Maritime University, Shanghai, 201306, China
2 Shanghai Ship and Shipping Research Institute Co., Ltd., Shanghai, 200135, China

* Corresponding Author: Zhuowei Wang. Email:

Computer Systems Science and Engineering 2025, 49, 377-399. https://doi.org/10.32604/csse.2025.059634

Received 13 October 2024; Accepted 28 February 2025; Issue published 09 April 2025

Abstract

Synthetic Aperture Radar (SAR) has become one of the most effective tools in ship detection. However, due to significant background interference, small targets, and challenges related to target scattering intensity in SAR images, current ship target detection faces serious issues of missed detections and false positives, and the network structures are overly complex. To address this issue, this paper proposes a lightweight model based on YOLOv8, named OD-YOLOv8. Firstly, we adopt a simplified neural network architecture, VanillaNet, to replace the backbone network, significantly reducing the number of parameters and computational complexity while ensuring accuracy. Secondly, we introduce a dynamic, multi-dimensional attention mechanism by designing the ODC2f module with ODConv to replace the original C2f module and using GSConv to replace two down-sampling convolutions to reduce the number of parameters. Then, to alleviate the issues of missed detections and false positives for small targets, we discard one of the original large target detection layers and add a detection layer specifically for small targets. Finally, based on a dynamic non-monotonic focusing mechanism, we employ the Wise-IoU (Intersection over Union) loss function to significantly improve detection accuracy. Experimental results on the HRSID dataset show that, compared to the original YOLOv8, OD-YOLOv8 improves mAP@0.5 and mAP@0.5–0.95 by 2.7% and 3.5%, respectively, while reducing the number of parameters and GFLOPs by 72.9% and 4.9%, respectively. Moreover, the model also performs exceptionally well on the SSDD dataset, with AP and AP50 increasing by 1.7% and 0.4%, respectively. OD-YOLOv8 achieves an excellent balance between model lightweightness and accuracy, making it highly valuable for end-to-end industrial deployment.

---

Keywords

---