MW-ACGAN: Generating Multiscale High-Resolution SAR Images for Ship Detection.

Area of Science:

• Remote Sensing
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning for high-resolution Synthetic Aperture Radar (SAR) ship detection is hindered by limited sample data.
• Existing methods struggle with low detection accuracy when training datasets are small.

Purpose of the Study:

• To develop a novel high-resolution SAR ship detection method for small sample scenarios.
• To improve the generation of realistic SAR ship images and enhance detection accuracy.

Main Methods:

• An improved Multiscale Wasserstein Auxiliary Classifier Generative Adversarial Networks (MW-ACGAN) was developed to generate high-resolution SAR ship images.
• Multi-scale loss terms and output layers were incorporated for generating diverse ship images.
• Generated images were combined with original data to train a Yolo v3 detection network.

Main Results:

• The MW-ACGAN successfully generated multi-scale and multi-class SAR ship images with high confidence (average score 0.91 for ResNet18).
• The Yolo v3 network trained on the composite dataset achieved a detection accuracy of 94%.
• This accuracy is significantly higher than training solely on the original SAR dataset.

Conclusions:

• The proposed MW-ACGAN effectively enhances SAR ship image generation for small sample datasets.
• Combining generated and real data substantially improves deep learning-based ship detection performance.
• The method optimizes the use of original data, leading to superior detection accuracy in high-resolution SAR imagery.