Conditional GANs for FL Sonar-to-Optical Image Transformation

Underwater robotics platforms are equipped with onboard cameras for navigation and mapping within deep-sea environments, either assisted by an operator in human-in-the-loop missions or automatically through real-time processing of live video. The impairment of vision by turbidity can be addressed by deploying the recent generation of 2-D multi-beam forward-looking imaging sonars, but this comes with limitations; e.g., sonar video data are not readily interpretable by an untrained human operator and cannot be effectively processed by the state-of-the-art computer methods. We propose a framework to translate underwater sonar imagery into intuitive optical views. By training a conditional generative adversarial network (Nvidia Pix2PixHD) on paired clear-water sonar and optical datasets, our model can synthesize high-quality optical images from sonar data captured in murky conditions. Evaluations across two distinct datasets—testing variables like training size, image resolution, field of view, color relevance, and cross-device transferability—demonstrate that our approach performs robustly in practical real-world scenarios.