Sonar image denoising based on clustering and Bayesian sparse coding

Area of Science:

• Marine technology
• Image processing
• Signal processing

Background:

• Side-scan sonar images (SSI) suffer from multiplicative speckle and additive noise, degrading quality and hindering interpretation.
• Effective denoising is crucial for accurate target recognition and scene analysis in sonar imagery.

Purpose of the Study:

• To develop an advanced denoising algorithm for SSI that addresses mixed noise.
• To enhance the preservation of structural details and target features in denoised images.

Main Methods:

• Integration of non-local similar block clustering with Bayesian sparse coding.
• Utilizing cross-scale structural features and noise statistics with an Equivalent Number of Looks (ENL) metric and improved K-means for patch classification.
• Employing a joint dictionary training strategy and Bayesian Orthogonal Matching Pursuit (BOMP) for sparse representation.

Main Results:

• The proposed algorithm effectively suppresses mixed noise (speckle and additive) in SSI.
• Demonstrated superior performance over classical methods in objective metrics (PSNR, SSIM) and visual quality.
• Significantly improved preservation of target edges and textures, even under severe noise conditions.

Conclusions:

• The proposed denoising algorithm offers a robust solution for enhancing SSI quality.
• It provides a valuable tool for improving target recognition and scene interpretation in marine acoustics.
• The method's ability to preserve structural details under noise is a key advantage.