Meso-structure segmentation of concrete CT image based on mask and regional convolution neural network

Abstract

The meso-damage of concrete structure has an important impact on the macro-damage. X-ray CT technology, as a non-destructive testing method, is one of great significance to the study of concrete meso-structure. However, since the traditional CT scanning image segmentation methods cannot effectively identify and segment structural meso-defects autonomously. In order to solve this problem, a new method for meso-crack segmentation and extraction of concrete CT scanning images based on regional convolution neural network (Mask R-CNN) is proposed and applied to segment the images obtained from the real-time CT test under static uniaxial compression. Results showed that the segmentation accuracy of voids and cracks in the CT images is 0.957 and 0.931 respectively, which verifies the effectiveness of the method that can be used to high-precision recognize voids and cracks from CT scanning images of concrete. Additionally, the segmented images of internal voids and cracks in concrete are adopted to accurately 3D reconstruction to investigate the meso-cracks propagation under real-time loading based on CT.

Keywords

• Concrete
• X-ray CT images
• Deep learning
• Mask R-CNN
• Segmentation