HFP-SAM: Hierarchical Frequency Prompted SAM for Efficient Marine Animal Segmentation.

Area of Science:

• Computer Vision
• Marine Biology
• Deep Learning

Background:

• Marine Animal Segmentation (MAS) faces challenges with long-distance modeling in complex environments.
• Existing deep learning methods for MAS often overlook fine-grained details and frequency information.
• The Segment Anything Model (SAM) shows promise but requires adaptation for specialized tasks like MAS.

Purpose of the Study:

• To develop a high-performance learning framework for Marine Animal Segmentation (MAS).
• To address the limitations of existing methods, particularly regarding detail perception and frequency analysis.
• To enhance the capabilities of the Segment Anything Model (SAM) for underwater imagery.

Main Methods:

• Proposed Hierarchical Frequency Prompted SAM (HFP-SAM) framework.
• Developed a Frequency Guided Adapter (FGA) to integrate marine scene information via frequency domain priors.
• Introduced Frequency-aware Point Selection (FPS) for generating informative prompts.
• Incorporated Full-View Mamba (FVM) for efficient spatial and channel contextual information extraction.

Main Results:

• HFP-SAM demonstrated superior performance in marine animal segmentation tasks.
• The framework successfully integrated frequency information and fine-grained details.
• Experiments on four public datasets validated the effectiveness of the proposed methods.

Conclusions:

• The HFP-SAM framework offers a significant advancement in marine animal segmentation.
• The integration of frequency analysis and enhanced prompting improves segmentation accuracy.
• The proposed approach provides a robust solution for identifying marine animals in challenging underwater conditions.