An explainable context-adaptive fusion and expert-in-the-loop evaluation framework for underwater sonar image...

Area of Science:

• Artificial Intelligence
• Marine Technology
• Computer Vision

Background:

• Sonar image interpretation is vital for underwater object detection.
• Conventional deep learning models struggle with sonar-specific features (e.g., acoustic shadows) and lack transparency.
• This opacity limits performance and user trust in AI-driven sonar analysis.

Purpose of the Study:

• To propose an explainable sonar image classification system addressing limitations of current deep learning models.
• To enhance the interpretability and trustworthiness of AI in sonar applications.
• To fuse specialized classifiers for improved sonar feature extraction and analysis.

Main Methods:

• Developed a novel Context-Adaptive Fusion Framework (CAFF) integrating Naive, ShadowNet, and HighlightNet classifiers via attention-based fusion.
• Incorporated explainability techniques: Gradient-weighted Class Activation Mapping (Grad-CAM), SHapley Additive exPlanations (SHAP), and Local Interpretable Model-agnostic Explanations (LIME).
• Implemented expert-in-the-loop evaluation using the Augmented QUality Assessment for eXplainability (AQUA-X) framework for validation and refinement.

Main Results:

• The CAFF effectively fuses sonar-specific features, outperforming conventional methods.
• Explainability techniques provided detailed insights into sonar-specific feature interpretation.
• Expert validation via AQUA-X confirmed the system's interpretability and identified optimal AI techniques for sonar analysis.

Conclusions:

• The proposed CAFF and AQUA-X framework significantly enhance the trustworthiness and transparency of AI for sonar image interpretation.
• This approach promotes reliable AI solutions for real-world underwater exploration and object identification.
• The study advances explainable AI (XAI) applications in specialized domains like marine acoustics.