Convolutional Neural Networks for 3D Inclusion Identification in Solid Media

In this paper we consider the problem of localization of large inclusions in a homogeneous solid medium. The general formulation corresponds to the problems typical for seismic or ultrasonic investigation. The problem is considered in a three-dimensional formulation. An acoustic model is used to describe the medium. A homogeneous background is defined in the whole region, in which 1‒2 large rigid inclusions are present. A synthetic data set obtained from numerical modeling is used. Convolutional neural networks are used to localize large inclusions. This work aims to test different variants of convolutional neural networks for processing the data of scanning heterogeneous solid medium in three-dimensional formulation. The result of this work is an estimation which network architectures can be applied for this problem. The work considers 7 different 2D conventional neural networks and different variations of 3D convolutional networks. All network architectures were tested on the same synthetic data. This work demonstrates that neural network using 3D UNet architecture with ResNet encoder trained on samples with one inclusion in the background media performs well on samples with two inclusions. The samples with two inclusions were not used in any way during the training. The network achieves accuracy 98.97% and dice score 0.833 in such testing. That is, 3D UNet architecture with ResNet encoder can be used for the problem of detecting inclusions in 3D media.