뉴럴 네트워크를 이용한 수중 전방 스캔 소나 이미지 영상 합성과 수중 물체 인식에의 응용

Abstract

This paper proposes a method to synthesize realistic underwater sonar images using a ray tracing and generative adversarial network. When a three-dimensional model of an object is given, a ray-tracing-based sonar simulator first calculates a base sonar image of the object. Then, the generative adversarial network translates the base images to the realistic sonar images by adding degradation effects such as noise and blurred edge. We evaluated the proposed method by comparing the synthesized images and the real sonar images captured in a water tank. The sonar images synthesized by the proposed method can be used to develop and verify the other sonar-based algorithms. We also verified the utility of the proposed method by applying the synthesized sonar image to deep-learning-based underwater object detection.