Denoising auto-encoder based image enhancement for high resolution sonar image

Abstract

A typical sonar image has a plenty of random noise compared to an optical image. Due to poor picture quality, there is a large restriction on recognizing any object. Pattern recognition is exceedingly difficult not only in computer image processing but even in human eyes. Numerous researchers have attempted to apply various types of average filters to sonar images, and have also removed noise by using multiple images in succession. However, each of the algorithms has a limitation in that the resolution of the image itself is degraded or the image of the object is difficult to remove noise. Finally, We performed sonar image noise reduction with the auto-encoder algorithm based on convolutional neural network, which as recently been attracting attention. With the algorithm, we obtained sonar images of superior quality with only a single continuous image. We simply learned a ton of sonar images in a neural network of auto-encoder structures, and then we could get the results by injecting the original sonar images. We verified the results of image enhancement using the acoustic lens based multibeam sonar images.