뉴럴 네트워크를 이용한 전방 스캔 소나 이미지에서의 누화 잡음 제거 방법

Abstract

This paper proposes a method to eliminate crosstalk noise in the multi-beam sonar images. In multi-beam sonar, crosstalk noise occurred easily near underwater objects because an acoustic wave reflected on the surface of the object can be mispercieved by the adjacent receivers in the sonar array. The crosstalk noise degrades the quality of sonar images and makes hard to distinguish underwater objects. The proposed method detects the region that the crosstalk noise occurs using convolutional neural network addressing that crosstalk noise has its own characteristic highlight pattern. Then, the proposed method eliminates the crosstalk noise by applying pixel adjustment in the detected region. Through this detect and removal steps, crosstalk noise can be removed efficiently from the multi-beam sonar images. The proposed method can be used as preprocessing of sonar images, and improve the reliability of sonar-based algorithms such as object detection and mapping.