SEGSID: A Semantic-Guided Framework for Sonar Image Despeckling.

Area of Science:

• Signal Processing
• Image Analysis
• Artificial Intelligence

Background:

• Sonar imagery quality is significantly degraded by speckle noise, necessitating effective despeckling techniques.
• Existing self-supervised despeckling methods, such as blind-spot networks (BSNs), struggle with speckle noise's spatial correlation and inherent information loss.

Purpose of the Study:

• To introduce SEGSID, a BSN-based semantic-guided framework for enhanced sonar image despeckling.
• To address the limitations of current methods in handling spatial noise correlation and information loss.

Main Methods:

• The SEGSID framework incorporates a Receptive Field Augmentation (RFA) module to extract local information while avoiding noise-correlated pixels.
• A Global Semantic Enhancement (GSE) module integrates global semantic information into local features to compensate for information loss.
• Knowledge distillation is employed to create a streamlined network for practical applications.

Main Results:

• SEGSID demonstrates superior performance in sonar image despeckling compared to traditional and state-of-the-art self-supervised methods.
• Experiments on three distinct sonar datasets validate the effectiveness of the proposed framework.
• The RFA and GSE modules successfully mitigate challenges posed by speckle noise correlation and information loss.

Conclusions:

• SEGSID offers a robust and efficient solution for sonar image despeckling, significantly improving image quality.
• The semantic-guided approach enhances the capability of BSNs to process complex sonar data.
• The publicly available implementation facilitates broader adoption and further research in sonar image processing.