Realistic Sonar Image Simulation Using Generative Adversarial Network

Abstract

Sonar sensors are widely utilized underwater because they can observe long-ranged objects and are tolerant to measurement conditions, such as turbidity and light conditions. However, sonar images have low quality and hard to collect, so development and application of sonar-based algorithms are difficult. This paper proposes a method to generate realistic sonar images or to segment real sonar image, to better utilize the sonar sensors. A simple sonar image simulator was implemented using a ray-tracing method. The simulator could calculate semantic information of real sonar images including properties of highlight, background, and shadow regions. Then, a generative adversarial network translated the simulated images into more realistic images or real sonar images into simulated-like images. The proposed method can be used to augment or pre-process sonar images.