Turbidity-Similarity Decoupling: Feature-Consistent Mutual Learning for Underwater Salient Object Detection.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Underwater salient object detection (USOD) is challenging due to image noise from water turbidity and low contrast between foreground and background.
• Existing methods struggle to effectively address both noise and similarity issues simultaneously.

Purpose of the Study:

• To propose a novel dual-model architecture for USOD that overcomes key detection challenges.
• To enhance the accuracy and robustness of underwater object detection systems.

Main Methods:

• Developed DenoisedNet using a separation-denoising-enhancement framework to suppress noise while preserving target features.
• Designed SearchNet with pseudo-label generation and layer-by-layer search for precise localization amidst low contrast.
• Implemented a feature-consistent mutual-learning strategy to enable collaborative learning between DenoisedNet and SearchNet.

Main Results:

• DenoisedNet and SearchNet demonstrated superior performance compared to existing methods on the USOD10K and USOD benchmarks.
• Achieved significant Mean Absolute Error (MAE) improvements: 4.52%/5.52% on USOD10K and 1.61%/8.94% on USOD.
• The mutual learning strategy effectively complemented the strengths of both models.

Conclusions:

• The proposed dual-model architecture effectively addresses the primary challenges in USOD.
• This approach offers a comprehensive solution for accurate and robust underwater salient object detection.
• The method achieves state-of-the-art performance, providing a valuable advancement in the field.