Fusion reconstruction of porous materials using 2D high- and 3D low-resolution images via latent representation regression generative adversarial network

Acquiring multi-scale microstructures is critical for understanding the physical and transport properties of porous materials. Due to the trade-off between imaging field of view (FoV) and resolution, it is challenges to simultaneously capture both large- and fine-scale characteristics of materials in a single imaging session. Multi-scale microstructure requires the fusion modeling of different resolution images. This paper proposes a fusion reconstruction framework based on the Latent Representation Regression Generative Adversarial Network (LRR-GAN), which attempts to construct multi-scale microstructure by fusing the structure features of low-resolution images and the detail features of high-resolution images. The framework comprises an improved Variational Autoencoder (VAE) and the LRR-GAN. The improved VAE is designed to decouple the detail features of high-resolution image, and the latent representation vector extracted by the pre-trained VAE encoder serves as a descriptor of fine-scale characteristics. The LRR-GAN generator is constructed to extract structure features of low-resolution images and fuses them with the detail features of high-resolution images. Further, a latent representation regression loss is presented to enable explicit measure of detail features, which guides the generator to generate multi-scale microstructures corresponding to the fine-scale characteristics of input high-resolution image. Experimental results demonstrate that the fusion results closely match the pore structure and flow characteristics of real porous rock, and fine-scale characteristics can be controlled based on the input high-resolution images. • Proposes a novel LRR-GAN for fusion reconstruction. • Improves a VAE to extract latent vectors as descriptors of fine-scale characteristic. • Introduces an LRR loss to constrain and improve fine-scale characteristic modeling. • Generates multi-scale structures based on specific I 2DHR and I 3DLS inputs. • Allows controllable reconstruction of fine-scale characteristic based on I 2DHR .