A SAR Ship Detection Method Based on Adversarial Training.

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Synthetic Aperture Radar (SAR) ship detection is crucial for maritime surveillance.
• Existing detectors struggle with generalization due to limited SAR image volumes.
• Traditional data augmentation methods offer minimal improvements in detection accuracy.

Purpose of the Study:

• To develop an adversarial training approach for generating synthetic SAR images to improve ship detection.
• To enhance the generalization ability and robustness of SAR ship detectors.
• To overcome the limitations of conventional data augmentation in SAR imaging.

Main Methods:

• Adversarial training is employed to generate new training samples by adding perturbations to SAR images.
• Batch normalization is separated for clean and perturbed samples to prevent performance degradation.
• K-step average perturbation and one-step gradient descent optimize the training process.
• Simultaneous perturbation and selection of high-loss samples improve adaptability to challenging scenarios.

Main Results:

• The proposed adversarial training method significantly improves SAR ship detection performance.
• Average Precision (AP) gains of 8%, 10%, and 17% were achieved on SSDD, SAR-Ship-Dataset, and AIR-SARShip, respectively.
• The method demonstrates superior results compared to traditional data augmentation techniques.

Conclusions:

• Adversarial training is an effective strategy for enhancing SAR ship detection by generating robust features.
• The proposed approach improves detector adaptability and performance across various datasets.
• This method offers a promising solution for improving the generalization ability of SAR ship detectors in diverse maritime environments.