Tianyi Guan

About

This project investigated fracture detection in low-contrast, high-noise CT imagery using two-dimensional cross-sections of cylindrical rock cores as a representative imaging problem. A multi-stage image-processing framework was developed to reduce background intensity variation, suppress high-frequency noise while preserving fracture boundaries, separate likely fracture regions from surrounding rock, and remove residual disconnected noise. The resulting workflow was designed to improve crack visibility and segmentation reliability under conditions where direct thresholding performs poorly.

Research Focus

To develop a robust image-processing workflow for identifying fracture structures in low-contrast CT cross-sections where noise, background inhomogeneity, and fuzzy boundaries limit the effectiveness of direct threshold segmentation.

Study Design

The processing pipeline consisted of four sequential stages. First, radial background intensity variation was corrected using the disk geometry of the core cross-section and Otsu-based separation of core and air regions. Second, non-local mean filtering was applied to suppress high-frequency noise while preserving structurally similar fracture boundaries. Third, intensity-based marking and linear stretching were used to separate likely fracture pixels from non-fractured regions. Finally, connected-region analysis removed residual isolated noise while retaining the dominant crack structure. The proposed workflow was compared visually with direct threshold segmentation and evaluated under different non-local mean denoising parameters.

Key Findings

Direct threshold segmentation failed to recover the full fracture structure reliably and was highly sensitive to threshold choice, particularly in noisy central regions. Background correction substantially reduced radial intensity bias, while non-local mean filtering improved noise suppression without eliminating major crack boundaries. Among the tested denoising settings, the lower smoothing parameter preserved edge structure more effectively, and the final connected-region filtering produced substantially cleaner fracture masks than direct thresholding.

Key Methods

Otsu thresholding; radial background correction; non-local mean filtering; low-contrast enhancement; adaptive intensity thresholding; linear gray-level stretching; connected-component analysis; graph-based denoising; CT image segmentation.

Keywords

Image processing; low-contrast imaging; non-local mean filtering; CT imaging; fracture detection; image denoising; background correction; threshold segmentation; connected components; rock-core imaging;