A diffusion-based super resolution model for enhancing sonar images.

Area of Science:

• Marine technology
• Artificial intelligence
• Image processing

Background:

• Modern sonar systems achieve centimeter resolution, but legacy datasets are limited.
• Practical constraints and older systems hinder resolution in existing sonar data.
• A unified dataset is needed for advanced machine learning applications.

Purpose of the Study:

• To upscale low-resolution sonar datasets using advanced AI.
• To ensure backward compatibility between legacy and high-resolution sonar data.
• To improve machine learning performance on unified sonar datasets.

Main Methods:

• Employed a single image super-resolution technique utilizing a conditioned diffusion model.
• Developed a mapping approach to upscale low-resolution sonar images to higher resolutions.
• Introduced two novel sonar-specific evaluation metrics grounded in acoustic physics and ATR utility.

Main Results:

• Upscaled images showed improved classification performance.
• The diffusion model approach did not increase the probability of false detections.
• Achieved a 7% higher probability of detection than bicubic interpolation, 6% over CNNs, and 2% over GANs.

Conclusions:

• Conditioned diffusion models effectively upscale sonar imagery for enhanced machine learning.
• The proposed method bridges resolution gaps in sonar datasets, enabling unified analysis.
• Novel metrics provide better evaluation of sonar image enhancement for target recognition.