Abstract

With the wide application of convolutional neural networks (CNNs), a variety of ship detection methods based on CNNs in synthetic aperture radar (SAR) images were proposed, but there are still two main challenges: (1) Ship detection requires high real-time performance, and a certain detection speed should be ensured while improving accuracy; (2) The diversity of ships in SAR images requires more powerful multi-scale detectors. To address these issues, a SAR ship detector called Duplicate Bilateral YOLO (DB-YOLO) is proposed in this paper, which is composed of a Feature Extraction Network (FEN), Duplicate Bilateral Feature Pyramid Network (DB-FPN) and Detection Network (DN). Firstly, a single-stage network is used to meet the need of real-time detection, and the cross stage partial (CSP) block is used to reduce the redundant parameters. Secondly, DB-FPN is designed to enhance the fusion of semantic and spatial information. In view of the ships in SAR image are mainly distributed with small-scale targets, the distribution of parameters and computation values between FEN and DB-FPN in different feature layers is redistributed to solve the multi-scale detection. Finally, the bounding boxes and confidence scores are given through the detection head of YOLO. In order to evaluate the effectiveness and robustness of DB-YOLO, comparative experiments with the other six state-of-the-art methods (Faster R-CNN, Cascade R-CNN, Libra R-CNN, FCOS, CenterNet and YOLOv5s) on two SAR ship datasets, i.e., SSDD and HRSID, are performed. The experimental results show that the AP50 of DB-YOLO reaches 97.8% on SSDD and 94.4% on HRSID, respectively. DB-YOLO meets the requirement of real-time detection (48.1 FPS) and is superior to other methods in the experiments.

Highlights

  • Automatic ship detection plays an important role in both civil and military fields, such as port management, fishery development supervision, and maritime rescue [1], etc

  • Traditional synthetic aperture radar (SAR) image ship detection methods are mostly based on concrete features and traditional image processing techniques, and are mainly classified through image segmentation, handcraft feature extraction and other methods for detection, such as the Constant False Alarm Rate (CFAR) [7], Histogram of Oriented Gradient (HOG) [8], Haar [9]

  • In order to ensure real-time detection, a single-stage detection method was selected as the basic network, and we reduced the parameters and calculation in a Feature Extraction Network (FEN) through a cross stage partial (CSP) [31] structure

Read more

Summary

Introduction

Automatic ship detection plays an important role in both civil and military fields, such as port management, fishery development supervision, and maritime rescue [1], etc. With the rapid increase in the data volume of SAR images to be processed, the requirements for accuracy and real-time performance of detection algorithms are increasing. The space-borne SAR system can achieve a high resolution of less than one meter [6], and the size differences of identifiable target ships increase, which poses a higher challenge to the multi-scale detection capability of the detection algorithms. Traditional SAR image ship detection methods are mostly based on concrete features and traditional image processing techniques, and are mainly classified through image segmentation, handcraft feature extraction and other methods for detection, such as the Constant False Alarm Rate (CFAR) [7], Histogram of Oriented Gradient (HOG) [8], Haar [9]

Methods
Results
Conclusion
Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call