InterPore2023

Semantic segmentation of rock images from multiple imaging methods using deep learning methods

23 May 2023, 10:30

1h 30m

Poster Presentation (MS15) Machine Learning and Big Data in Porous Media Poster

Speaker

Hongkyu Yoon (Sandia National Laboratories)

Description

Over the past decade non-destructive imaging methods for materials have been increasingly advanced. Two most notable imaging methods include X-ray computed microtomography (μCT) that can image materials at sub-micron scale to millimeter scale resolutions and focused ion beam-scanning electron microscopy (FIB-SEM) that can image at a nano-meter scale. Hence, the segmentation of images obtained from different imaging techniques is a critical step towards quantitatively describing various features of geomaterials over a range of scales. In this work we evaluate various deep learning methods (e.g., U-Net, Attention U-Net, Efficient net, transformer, VGG16, ResNet, and MultiResUnet) to segment both μCT and FIB-SEM images. Four independent datasets including sandstone, carbonate chalks, and shale are evaluated. Each of these datasets is composed of three-dimensional image stacks and corresponding ground truth segmentation labels obtained using various traditional image processing techniques. Our preliminary results indicate that deep learning architectures can successfully be applied to the task of semantic segmentation for individual dataset with frequency weighted accuracy between 94% and 99% (on testing data) and can perform better than manual segmentation to recover the natural morphology of original images. However, performance is significantly deteriorated by ~ 10-30% in accuracy when mixed images from different imaging methods are used as training data. Here, we will demonstrate the improvement of semantic segmentation of multiple rock images from both μCT and FIB-SEM through transfer learning of transformers and other deep learning methods.
SNL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525.