Importance of appropriate segmentation in pore structure analysis of coral reef limestone from CT images

Abstract

This paper conducts a critical evaluation of the applicability of traditional methods and neural network-based methods in separating pore space from the matrix on computed tomography (CT) scanning images of two types of coral reef limestone specimens with coarse and fine pores. By comparing with the manually labelled “ground truth,” it is shown that the Otsu method and other traditional methods have good segmentation accuracy and high efficiency when subjected to specific image preprocessing schemes, but their robustness is poor. UNet neural networks with a mixed contrast training dataset have high accuracy and strong robustness, but they are complex and computationally expensive. The three-dimensional reconstruction and analysis of the spatial distribution of porosity shows that the coral reef limestone specimen with coarse pores has a clear pore structure, while the rock matrix and pores are interlaced in the specimen with fine pores. Last, the statistical analysis of porosity and the pore distributions prove that the segmentation accuracy has a significant influence on the subsequent pore analysis. The current study highlights the importance of using appropriate image segmentation schemes and provides useful guidance for future interpretation of pore structure features of coral reef limestone based on CT scanned images.