Ship Detection Using Edge-Based Segmentation and Histogram of Oriented Gradient with Ship Size Ratio

Hyukmin Eum, Jaeyun Bae, Changyong Yoon, Euntai KimInt. J. Fuzzy Log. Intell. Syst 2015 · DOI

Abstract

In this paper, a ship detection method is proposed; this method uses edge-based segmentation and histogram of oriented gradient (HOG) with the ship size ratio. The proposed method can prevent a marine collision accident by detecting ships at close range. Furthermore, unlike radar, the method can detect ships that have small size and absorb radio waves because it involves the use of a vision-based system. This system performs three operations. First, the foreground is separated from the background and candidates are detected using Sobel edge detection and morphological operations in the edge-based segmentation part. Second, features are extracted by employing HOG descriptors with the ship size ratio from the detected candidate. Finally, a support vector machine (SVM) verifies whether the candidates are ships. The performance of these methods is demonstrated by comparing their results with the results of other segmentation methods using eight-fold cross validation for the experimental results.

Introduction

Marine collisions remain a significant threat despite modern navigation technology. Two primary factors contribute to these accidents: human limitations and radar blind spots. Research indicates that human operators can only maintain peak concentration for approximately 20 minutes, and their physical field of view is inherently restricted. Meanwhile, traditional radar systems often struggle to detect small vessels, wooden boats, or ships constructed with radio-absorbent materials. To mitigate these risks, especially in crowded harbors or narrow channels, there is a pressing need for an automated, vision-based system capable of identifying diverse vessel types in real-time.

Methodology

The proposed system utilizes a three-stage pipeline to achieve robust ship detection, moving from raw input to final classification.

Flowchart of the proposed ship detection system.

1. Edge-Based Segmentation

Rather than relying on color or intensity—which can be unreliable when a ship's hull color mimics the surrounding water—the system employs Sobel edge detection. This is followed by morphological operations, including dilation, closing, and opening, to bridge gaps in edges and filter out noise. This process effectively isolates potential ship "candidates" from the dynamic maritime background.

Candidate detection result after morphological operations.

2. Feature Extraction (HOG + Ship Size Ratio)

Features are extracted using the standard Histogram of Oriented Gradients (HOG) descriptor to capture local shape information. A novel contribution of this work is the addition of a Ship Size Ratio to the feature vector. Ships typically adhere to specific geometric proportions (e.g., a length-to-width ratio not exceeding 1:3). By normalizing and appending this physical constraint, the model can better distinguish actual vessels from non-ship objects like waves or coastal buildings. The authors emphasize that the ratio must be scaled appropriately so its deviation does not overwhelm the HOG values during training.

3. Classification

A weighted nu-SVM with a Radial Basis Function (RBF) kernel is used to verify the candidates. The RBF kernel was selected after comparative testing against Linear, Polynomial, and Sigmoid kernels for its superior performance in this domain.

Dataset and Evaluation

The system was evaluated using a custom dataset of approximately 11,000 image frames at 320 × 240 resolution. The dataset encompasses eight distinct ship types, ranging from 70m passenger ships to 3m outboard motorboats, ensuring the model's robustness across various scales.

Results and Discussion

The experimental results highlight the effectiveness of integrating geometric constraints with visual descriptors.

  • Segmentation Performance: The Sobel mask outperformed other edge detection methods (such as Prewitt, Robert, and LoG) with a precision of 0.4046 for candidate detection. It proved particularly effective at maintaining ship boundaries while ignoring "wake noise" produced by moving vessels.
  • Detection Accuracy: The complete system achieved a peak accuracy of 93.6% using eight-fold cross-validation.
  • Impact of Size Ratio: The inclusion of the ship size ratio provided a measurable boost in both accuracy and precision compared to standard HOG. This suggests that visual descriptors alone are insufficient; incorporating physical dimensions significantly reduces false positives.

Conclusion

This vision-based approach offers a promising supplement to traditional radar and human lookout. By combining edge-based segmentation with HOG descriptors and ship size ratios, the system achieves high accuracy across diverse vessel types. Future work will focus on optimizing the system for high-resolution video to ensure real-time performance and improving environmental robustness against adverse weather conditions like fog and heavy rain.

Summary updated

BibTeX
@article{eum2015ship,
  author = {Hyukmin Eum and Jaeyun Bae and Changyong Yoon and Euntai Kim},
  title = {{Ship Detection Using Edge-Based Segmentation and Histogram of Oriented Gradient with Ship Size Ratio}},
  journal = {Int. J. Fuzzy Log. Intell. Syst},
  year = {2015},
  doi = {10.5391/ijfis.2015.15.4.251},
}
Figure