Real-Time Long-Distance Ship Detection Architecture Based on YOLOv8

Long-distance detection of maritime ships is pivotal for the development of intelligent collision avoidance systems. Despite significant advancements in target detection achieved through deep learning, the identification of long-distance ships poses a substantial challenge due to their small pixel s...

Full description

Saved in:
Bibliographic Details
Published in:IEEE access 2024, Vol.12, p.116086-116104
Main Authors: Gong, Yanfeng, Chen, Zihao, Deng, Wen, Tan, Jiawan, Li, Yabin
Format: Article
Language:English
Subjects:
Accuracy
Algorithms
Collision avoidance
Datasets
Deep learning
Feature extraction
Long-distance ship
Machine learning
Marine vehicles
Maritime communications
Neural networks
Object recognition
Real time
Real-time systems
Ships
small object detection
Target detection
Transformers
YOLO
YOLOv8
Citations: Items that this one cites
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Long-distance detection of maritime ships is pivotal for the development of intelligent collision avoidance systems. Despite significant advancements in target detection achieved through deep learning, the identification of long-distance ships poses a substantial challenge due to their small pixel size in images. Consequently, the recognition of long-distance ships essentially amounts to small object detection. In response to these challenges in small object detection, this paper proposes Ship-YOLOv8, a modified architecture derived from You Only Look Once version 8 (YOLOv8). First, we developed the C-Bottleneck Transformer neural network (C-BoTNet), which is integrated at the end of the backbone, to enhance the global receptive field and facilitate feature fusion. Additionally, we incorporated shallow features with deep features and introduced a dedicated detection layer for small objects into the original structure. Furthermore, we optimized the C2f in the neck using the cross stage partial network (VoVGSCSP) based on GSConv. Finally, we conducted optimization using the Wise-IoU loss function. Extensive experiments conducted on a self-created dataset of long-distance ships demonstrate the remarkable capabilities of Ship-YOLOv8. The proposed method achieves an AP0.5 of 91.8%, significantly outperforming YOLOv8's AP0.5 of 70.6%. Moreover, our method attains a detection speed of 4.8 ms per image during inference, showcasing its efficiency in real-time applications. To validate the algorithm's broad applicability, comparative experiments were conducted on a public maritime dataset SeaShips. Ship-YOLOv8 achieved an AP0.5 score of 99.3%, surpassing YOLOv8's 98.6%. Code is available at https://github.com/zihaohao123/Ship-YOLOv8 .
ISSN:2169-3536
2169-3536
DOI:10.1109/ACCESS.2024.3445154