Application of a regressive neural network with autoencoder for monochromatic images in ultrasound tomography

Abstract

The article presents a novel approach to ultrasound tomography in industrial applications. In order to visualize the interior of a tank (reactor) filled with tap water, a single neural network enhanced with autoencoder was used. A novelty is the use of an autoencoder to improve the quality of the measurement vector. Thanks to the use of the autoencoder for denoising the input measurements in connection with the appropriately adapted neural network, the quality of the output image was improved. A robust algorithm was developed that properly reconstructs hidden objects in monochrome images with high efficiency.