Sonar Image Translation Using Generative Adversarial Network for Underwater Object Recognition

Sonar sensor is widely used for underwater object recognition. However, acquiring reference sonar images for each target object is high-cost and time-consuming. Sonar image simulators can generate reference sonar images with small computation, but the simulated images are different with actual sonar images captured in the field. This paper proposes a method to translate actual sonar images to simulated-like images using a generative adversarial network. We trained the network with images captured by the indoor water tank test. The trained neural network can generate simulator-like images from given actual sonar images. Further, we can recognize the target object using template matching between the translated image and the reference images simulating the target object.