Fine-Grained Ship Recognition from the Horizontal View Based on Domain Adaptation.

Area of Science:

• Computer Vision
• Maritime Technology
• Machine Learning

Background:

• Ship recognition is crucial for maritime safety and management.
• Existing methods often rely on SAR or space-borne optical images, with limited fine-grained classification in visible spectrum.
• Visible image-based ship classification faces challenges with varying angles and domain gaps.

Purpose of the Study:

• To develop a robust fine-grained ship classification system using visible images.
• To address the domain gap between simulated and real-world ship images.
• To improve classification accuracy for ships at different viewing angles.

Main Methods:

• Constructed a fine-grained ship dataset with real and simulated images across five categories.
• Employed style transfer to minimize the visual differences between simulated and real images.
• Utilized a domain adaptation network with local maximum mean discrepancy (LMMD) for domain alignment.
• Integrated a Vision Transformer (ViT) for fine-grained feature extraction and a fully connected layer for classification.

Main Results:

• Achieved an overall accuracy of 96.0% on the fine-grained ship dataset.
• Attained a mean average precision (mAP) of 87.5% for detection and classification tasks.
• Demonstrated the feasibility of using computer-generated simulation images for auxiliary training.

Conclusions:

• The proposed network effectively performs fine-grained ship classification in visible images.
• Domain adaptation techniques successfully bridge the gap between simulated and real image domains.
• Simulated imagery can be a valuable resource for enhancing ship recognition model training.