Class-incremental few-shot underwater object detection framework.

Area of Science:

• Marine Biology
• Computer Vision
• Artificial Intelligence

Background:

• Underwater object detection is crucial for marine science but hindered by poor image quality, species diversity, and limited data.
• Existing methods face challenges in simultaneously addressing these issues, impacting ecological monitoring and exploration.
• The scarcity of labeled underwater imagery limits the development of robust detection models.

Purpose of the Study:

• To propose a novel class-incremental few-shot underwater object detection framework (CIFS-UD).
• To enhance underwater image quality and feature representation for improved detection.
• To mitigate catastrophic forgetting in incremental learning scenarios for underwater object detection.

Main Methods:

• An end-to-end multi-task architecture integrating deep inception and channel-wise attention (DICAM) for image enhancement and object detection.
• A prototype-based multi-scale dual attention module (ProMSDA) to optimize feature correlations and object region representation.
• A dynamic boundary-aware prototype collaboration optimization (DB-PCO) strategy to ensure class compactness and separability.

Main Results:

• The proposed CIFS-UD framework effectively adapts to new underwater object classes.
• Significant improvements in class-incremental few-shot underwater object detection performance were observed.
• Experiments on Brackish, Trashcan, and RUOD datasets validated the framework's efficacy.

Conclusions:

• CIFS-UD offers a robust solution for few-shot, class-incremental underwater object detection.
• The framework successfully balances learning new classes with retaining knowledge of old ones.
• This advancement supports more effective marine exploration and ecological monitoring through improved automated analysis.