Image Noise Reduction based on Generative Adversarial Networks by Underwater Acoustic Cameras

Abstract

Sensing the surrounding environment with acoustic cameras is becoming increasingly important for unmanned robot activities in extreme underwater environments that are inaccessible to humans. However, in turbid underwater environments, noise that cannot be reduced by conventional methods is generated in the acquired acoustic images due to diffuse reflection from the ground. In this study, we attempt to reduce noise in acoustic images using generative adversarial networks (GAN), an AI-based image generation method. Experimental results show that contrastive learning for unpaired image-to-image translation (CUT), one of the extension technologies of GAN, can reduce noise in acoustic images.