A Feature Optimization Approach Based on Inter-Class and Intra-Class Distance for Ship Type Classification.

Area of Science:

• Marine engineering
• Artificial intelligence
• Signal processing

Background:

• Deep learning excels at ship type classification, but relies on time-frequency (TF) features whose discriminative power is not fully exploited.
• Existing methods often struggle with subtle differences between ship types, leading to classification errors.

Discussion:

• A novel feature optimization method is proposed, minimizing an objective function to increase inter-class and decrease intra-class feature distances.
• This method learns class-specific centers, pulling samples of the same class closer to their respective centers.
• The optimization maximizes discriminative information, particularly for easily confused ship types.

Key Insights:

• The proposed method effectively enhances the discriminative power of features for ship type classification.
• Learned feature representations show improved separation between different ship classes.
• Outperforms conventional TF features on a real-world dataset.

Outlook:

• Potential for broader application in other classification tasks requiring fine-grained feature discrimination.
• Further research could explore adaptive learning rates and different objective functions.
• Integration with advanced deep learning architectures may yield further performance gains.