Related Experiment Video
Updated: Jun 7, 2025
03:31
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
471
Two-stage ship detection at long distances based on deep learning and slicing technique
Yanfeng Gong1, Zihao Chen1, Jiawan Tan1
1School of Shipping and Naval Architecture, Chongqing Jiaotong University, Chongqing, China.
Plos One
|November 19, 2024
Summary
This study introduces a two-stage deep learning approach for robust long-distance ship detection. The method enhances accuracy by first identifying the sea-sky line region and then analyzing ship patches, improving intelligent ship perception.
Area of Science:
- Computer Vision
- Artificial Intelligence
- Maritime Technology
Background:
- Accurate ship detection over long distances is vital for intelligent ship visual perception.
- Traditional image processing methods lack robustness for small, distant ships.
- Deep learning methods struggle with feature extraction from low-pixel, distant ships.
Purpose of the Study:
- To develop a robust and accurate method for detecting ships over long distances.
- To combine traditional and deep learning techniques for improved ship detection.
- To enhance the capabilities of intelligent ships in maritime surveillance.
Main Methods:
- A two-stage object detection framework is proposed.
- Stage 1: Detects the sea-sky line (SSL) region using YOLOv8 to identify potential ship areas.
- Stage 2: Applies another YOLOv8 model to sliced patches within the SSL region for precise ship detection.
Main Results:
- Achieved 85% average precision (AP50).
- Demonstrated a fast detection speed of 75 ms per image (1080x640 resolution).
- The two-stage approach effectively addresses challenges of detecting small, distant ships.
Conclusions:
- The proposed two-stage method significantly improves long-distance ship detection accuracy and efficiency.
- This approach offers a robust solution for intelligent ship visual perception systems.
- The integration of SSL detection and patch-based analysis enhances feature extraction for small objects.
