Extraction of fractures in shale CT images using improved U-Net

Abstract

Accurate extraction of pores and fractures is a prerequisite for constructing digital rocks for physical property simulation and microstructural response analysis. However, fractures in CT images are similar in grayscale to the rock matrix, and traditional algorithms have difficulty to achieve accurate segmentation results. In this study, a dataset containing multiscale fracture information was constructed, and a U-Net semantic segmentation model with a scSE attention mechanism was used to classify shale CT images at the pixel level and compare the results with traditional methods. The results showed that the CLAHE algorithm effectively removed noise and enhanced the fracture information in the dark parts, which is beneficial for further fracture extraction. The Canny edge detection algorithm had significant false positives and failed to recognize the internal information of the fractures. The Otsu algorithm only extracted fractures with a significant difference from the background and was not sensitive enough for fine fractures. The MEF algorithm enhanced the edge information of the fractures and was also sensitive to fine fractures, but it overestimated the aperture of the fractures. The U-Net was able to identify almost all fractures with good continuity, with an MIou and Recall of 0.80 and 0.82, respectively. As the image resolution increases, more fine fracture information can be extracted.

Keywords

• CT slices
• Fracture segmentation
• Shale
• U-Net
• Deep learning