Rock CT Image Fracture Segmentation Based on Convolutional Neural Networks

Image-based automatic fracture extraction methods have many practical applications in geological and engineering. Fracture identification and quantitative characterization require the means of interpreting and statistically analyzing image data. In comparison to traditional digital image processing methods, supervised semantic segmentation methods based on Convolutional Neural Networks (CNN) offer distinct advantages in extracting fractures from CT images. The study analyzes the characteristics of fracture areas in CT images and compares the results obtained through traditional threshold segmentation methods with those achieved using deep learning techniques. An integrated approach combining interactive image segmentation was proposed in this study to extract the fractures, in which deep learning methods were to determine the fracture areas, as well as morphological operation and threshold segmentation methods were adopted to extract the fractures from the small target areas. The implementation of this method significantly enhances the efficiency of extraction results compared to manual fracture extractions. This study selected CT images of igneous rock with a resolution of 0.7 um as the research object. DeepLab V3 + and UNet3 + network models in the PaddleSeg framework were used for the deep learning process.