Marine Saliency Segmenter: Object-Focused Conditional Diffusion With Region-Level Semantic Knowledge Distillation

Area of Science:

• Computer Vision
• Marine Biology
• Artificial Intelligence

Background:

• Marine Saliency Segmentation (MSS) is crucial for underwater vision tasks.
• Existing methods struggle with imprecise boundaries due to underwater environmental challenges like low contrast and color distortion.
• Diffusion models show promise but haven't fully leveraged contextual semantics for marine object feature learning.

Purpose of the Study:

• To develop a novel marine saliency segmentation method, DiffMSS, leveraging diffusion models and semantic knowledge distillation.
• To enhance feature learning for region-level salient objects in marine environments.
• To improve the accuracy and structural fidelity of marine instance segmentation.

Main Methods:

• Proposed DiffMSS, a marine saliency segmenter based on diffusion models.
• Introduced Word-level Semantic Saliency Extraction to identify salient terms from captions via region-word similarity.
• Utilized semantic knowledge distillation to generate diffusion conditions for the Conditional Feature Learning Network.
• Employed an Object-Focused Conditional Diffusion module and Consensus Deterministic Sampling for fine-grained segmentation masks and reduced mis-segmentations.

Main Results:

• DiffMSS demonstrated superior performance compared to state-of-the-art methods in quantitative and qualitative evaluations.
• The method effectively addresses challenges of imprecise boundaries in marine environments.
• Achieved fine-grained segmentation masks with enhanced structural fidelity.

Conclusions:

• DiffMSS offers a significant advancement in marine saliency segmentation.
• The integration of diffusion models with semantic knowledge distillation proves effective for underwater vision tasks.
• The proposed approach enhances the accuracy and reliability of marine exploration systems.