Microstructure synthesis using style-based generative adversarial networks

This work considers the usage of StyleGAN architecture for the task of microstructure synthesis. The task is the following: Given number of samples of structure we try to generate similar samples at the same time preserving its properties. Since the considered architecture is not able to produce samples of sizes larger than the training images, we propose to use image quilting to merge fixed-sized samples. One of the key features of the considered architecture is that it uses multiple image resolutions. We also investigate the necessity of such an approach.

Figure 1

Examples of original and reconstructed images from Ref. [1] [(a) and (b)] and Ref. [2] (c).

Figure 2

An example of samples drawn from the given structure.

Figure 3

A three-layer fully connected neural network (input is usually not regarded as a layer).

Figure 4

Deconvolution with stride 2 and kernel $3\times 3$, (green pixels) pixels of the image, white is padding, and (blue) kernel of convolution.

Figure 5

Generator scheme: (a) generator as a sequence of blocks and (b) operations in blocks: A, affine transform; B, per-channel scaling.

Figure 6

Discriminator scheme: (a) discriminator as a sequence of blocks and (b) operations in blocks.

Figure 7

Schematic result of image quilting.

Figure 8

Training loss for Alporas images.

Figure 9

Visual comparison of obtained results for Alporas: (a) original image of alporas, (b) binarized image with artifacts, (c) binarized image with filling of the holes, (d) result of StyleGAN, and (e) binarized StyleGAN result.

Figure 10

Training losses: (a) on the Berea dataset and (b) on the Ketton dataset.

Figure 11

Visual comparison of the results for Berea: (a) original image, (b) result of StyleGAN+quilting, and (c) result of PorousMediaGAN.

Figure 12

Visual comparison of the results for Ketton: (a) original image, (b) result of StyleGAN+quilting, and (c) result of PorousMediaGAN.

Figure 13

Two-point correlations: (a) for Berea and (b) for Ketton.

Figure 14

Minkowski functionals distributions for Berea: (a) distribution of area density, (b) distribution of perimeter density, (c) distribution of Euler2d density, (d) distribution of the number of pores, and (e) distribution of the number of solid areas.

Figure 15

Minkowski functionals distributions for Ketton: (a) distribution of area density, (b) distribution of perimeter density, (c) distribution of Euler2d density, (d) distribution of the number of pores, and (e) distribution of the number of solid areas.

Figure 16

The process of generating new structures.