Preparation of a Training Dataset for Reconstructing Three-dimensional Structures Using a Single Two-dimensional Image

Abstract

The construction of three-dimensional porous media structures using generative adversarial network models is currently a hot research topic. When using this method for 3D reconstruction, a rich training data is required, and the quality of the database seriously affects the quality of the final 3D reconstruction results. In practical situations, sometimes only a limited number of two-dimensional slice images can be obtained. Therefore, how to use the limited two-dimensional slice images to construct the training data required for generating adversarial network models has important practical significance and research value. Based on this, this paper focuses on how to prepare a training data for reconstructing three-dimensional structures in the presence of only a single two-dimensional image. There are two stages used in our method. In the first stage, digital techniques such as cropping, bilinear interpolation, corrosion, and dilation are applied to the obtained images to enrich our database. In the second stage, the optimal proportion of images obtained from each part was found through extensive experiments, resulting in the best performance of reconstructing the three-dimensional structure. In the final part of this article, a set of core samples were used to validate the above method. The experiment showed that the proposed method of constructing a 3D reconstruction training data using a single 2D image was effective.