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

Abstract

Improved hardware and processing techniques such as synthetic aperture sonar have led to imaging sonar with centimeter resolution. However, practical limitations and old systems limit the resolution in modern and legacy datasets. This study proposes using single image super resolution based on a conditioned diffusion model to map between images at different resolutions. This approach focuses on upscaling legacy, low-resolution sonar datasets to enable backward compatibility with newer, high-resolution datasets, thus creating a unified dataset for machine learning applications. The study demonstrates improved performance for classifying upscaled images without increasing the probability of false detection. The increased probability of detection was 7% compared to bicubic interpolation, 6% compared to convolutional neural networks, and 2% compared to generative adversarial networks. The study also proposes two sonar specific evaluation metrics based on acoustic physics and utility to automatic target recognition.