Convolutional Neural Network Based Resolution Enhancement of Underwater Sonar Image Without Losing Working Range of Sonar Sensors

Abstract

In underwater environment, sonar sensors have the advantage of being able to shoot images in turbid environment and having long working range. However, images taken with sonar sensor are difficult to recognize because of their low resolution. This paper proposes neural network based efficient resolution enhancement method in sonar images. We built convolutional neural network composed of 23 convolutional layers and 18 ResNet blocks, and trained the network with actual and denoised underwater sonar images. As a result, high resolution images can be restored from manually lowered resolution images, recording higher PSNR compared to interpolation algorithms. The proposed method can increase resolution of noisy, low-resolution sonar images without loss in working range.