Realistic Sonar Image Simulation Using Deep Learning for Underwater Object Detection

Abstract

This paper proposes a method that synthesizes realistic sonar images using a Generative Adversarial Network (GAN). A ray-tracing-based sonar simulator first calculates semantic information of a viewed scene, and the GAN-based style transfer algorithm then generates realistic sonar images from the simulated images. We evaluated the method by measuring the similarity between the generated realistic images and real sonar images for several objects. We applied the proposed method to deep learning-based object detection, which is necessary to automate underwater tasks such as shipwreck investigation, mine removal, and landmark-based navigation. The detection results showed that the proposed method could generate images realistic enough to be used as training images of target objects. The proposed method can synthesize realistic training images of various angles and circumstances without sea trials, making the object detection straightforward and robust. The proposed method of generating realistic sonar images can be applied to other sonar-image-based algorithms as well as to object detection.