Multi-Level Cross-Modal Interactive-Network-Based Semi-Supervised Multi-Modal Ship Classification.

Area of Science:

• Computer Vision
• Machine Learning
• Marine Technology

Background:

• Ship image classification is crucial for marine applications.
• Multi-modal approaches combining visible and infrared images enhance classification by leveraging complementary information.
• Current methods often fail to capture multi-level cross-modal correlations and require extensive labeled data.

Purpose of the Study:

• To develop a novel semi-supervised multi-modal ship classification approach.
• To address the limitations of existing methods in capturing multi-level cross-modal correlations and reducing reliance on labeled data.

Main Methods:

• A multi-level cross-modal interactive network is proposed to learn local and global feature correlations between different modalities.
• A semi-supervised contrastive learning strategy is employed to optimize the network using both labeled and unlabeled data.
• The strategy enforces intra-class consistency using supervision signals from unlabeled samples and prior labels.

Main Results:

• The proposed approach achieves state-of-the-art performance in semi-supervised ship classification.
• Experiments on public datasets validate the effectiveness of the multi-level cross-modal interaction and semi-supervised learning strategy.
• The method successfully captures comprehensive complementary information for improved classification.

Conclusions:

• The novel semi-supervised approach significantly enhances ship image classification accuracy.
• The developed multi-level cross-modal interactive network and contrastive learning strategy effectively utilize unlabeled data.
• This work offers a promising direction for efficient multi-modal ship classification in marine fields.