Prompt Then Refine: Prompt-Free SAM-Enhanced Collaborative Learning Network for Detecting Salient Objects in...

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Underwater salient object detection (USOD) faces challenges like poor lighting and blur.
• Traditional methods struggle with USOD due to these environmental factors.
• The Segment Anything Model (SAM) shows promise but requires prompt labels, which are scarce in USOD datasets.

Purpose of the Study:

• To develop a prompt-free network for effective USOD using SAM.
• To enhance SAM's performance in underwater conditions without manual annotations.
• To improve the accuracy and robustness of salient object detection in underwater imagery.

Main Methods:

• Proposed SAM-CLNet, a collaborative learning network integrating SAM, a mask prompt generator (MPG), and a region-aware attention collaborative learning loss (RCL).
• Utilized MPG to generate pseudo-mask prompts for SAM, addressing the lack of manual labels.
• Implemented a cyclic feedback mechanism where RCL refines MPG using SAM's predictions, creating mutual improvement.
• Introduced a U-Adapter for SAM adaptation to underwater imagery and a frequency cross-attention fusion module in MPG for RGB-depth integration.

Main Results:

• SAM-CLNet demonstrated superior performance compared to existing methods on USOD10K and USOD datasets.
• The network showed effective generalization across five public salient object detection (SOD) benchmarks.
• The prompt-free approach successfully compensated for the absence of manual labels, enhancing detection accuracy.

Conclusions:

• SAM-CLNet offers a robust and effective solution for prompt-free underwater salient object detection.
• The collaborative learning framework and specialized modules significantly improve detection in challenging underwater scenes.
• The proposed method advances the capabilities of SAM for specialized domain applications like USOD.