Multi-phase segmentation methods for micro-tomographic images based on deep learning

Abstract

Micro-tomography enables the acquisition of three-dimensional images of the internal microstructure of coal, which provides essential information for reservoir evaluation, mining planning, and coalbed methane extraction. However, the intricate issue of multi-phase segmentation in microscopic tomographic images has significantly hindered the efficient advancement of subsequent research endeavors. Traditional segmentation methodologies, which necessitate manual labor, are not only time-consuming and arduous but also inherently prone to errors, thereby failing to align with the contemporary industrial demands for high precision and efficiency. Therefore, the efficient and accurate segmentation of these complex micro-tomographic images, particularly the achievement of multi-phase segmentation, is of urgent necessity. To accurately and swiftly establish digital core images of multi-component coal, in this paper, we propose a novel multi-phase segmentation system for micro-tomography images, leveraging deep learning algorithms Utilizing coal CT images as the primary dataset and incorporating interactively threshold-segmented images as labels, we innovatively employ the U-Net model for automated segmentation training. Through rigorous experimental validation and analysis, the trained U-Net model demonstrates exceptional performance in mineral content identification, morphological feature extraction, and spatial structure analysis. When compared to traditional methods, the error rate is markedly decreased, and segmentation efficiency is enhanced by an order of magnitude. This innovative approach transcends the constraints of traditional manual segmentation. Leveraging the robust feature-learning capabilities of deep neural networks, it facilitates intelligent and rapid conversion from raw grayscale images to multi-component images, substantially improving segmentation accuracy and efficiency. This technique addresses the technological gap in swiftly and precisely constructing multi-component digital core images, offering a novel technical pathway for detailed reservoir evaluation, scientific mining plan development, and efficient coalbed methane extraction.