Multi-scale feature fusion keypoint detection network for ship draft line localization.

Area of Science:

• Maritime technology
• Computer vision
• Deep learning

Background:

• Accurate ship draft line detection is vital for maritime transaction fairness and navigational safety.
• Current deep learning methods often use segmentation, leading to high computational costs and challenges with environmental variations.
• Existing methods struggle with varying lighting and hull colors, impacting reliability.

Purpose of the Study:

• To develop a precise and efficient method for ship draft line detection.
• To overcome the limitations of existing segmentation-based deep learning approaches.
• To improve the adaptability of draft line detection models to diverse environmental conditions.

Main Methods:

• Proposed a Multi-scale Feature Fusion Keypoint Detection Network (MFFKD).
• Integrated Dilated Residual-Channel Recalibration Module (DR-CRM) blocks for multi-scale feature extraction.
• Employed Feature Enhancement Extraction Modules (FEEM) and Multi-scale Feature Weighted Integration (MFWI) for feature fusion.
• Utilized a keypoint prediction task head and character recognition for precise waterline readings.
• Implemented a dual-phase training strategy: pre-training and fine-tuning.

Main Results:

• The MFFKD method demonstrated superior accuracy compared to baseline models in waterline detection.
• Achieved significantly faster execution speeds than advanced segmentation-based approaches.
• The keypoint detection approach proved effective in diverse environmental conditions.

Conclusions:

• Integrating keypoint detection with dual-phase training offers an effective solution for ship waterline detection.
• The proposed MFFKD network provides a more accurate and computationally efficient alternative.
• The method enhances reliability in real-world maritime scenarios.