Abstract: Four 20 mm cubic 3D braided carbon/carbon (C/C) composite specimens were scanned by micro X-ray computed tomography (XCT) to obtain internal microstructure images with a voxel resolution of 18.27 μm. A deep learning based semantic segmentation algorithm was then used to train a large number of 2D XCT images to achieve intelligent identification and segmentation of rods, fiber bundles, matrix, pores, delamination and cracks of these specimens. The results show that: (1) The XCT scanning can characterize the distribution and morphology of the above components and defects with high resolutions, and the dominant defect is delamination between adjacent fiber bundle layers; (2) Since the grey values in the CT images of all micro components of C/C composites are very close, it is impossible for the traditional threshold segmentation method to segment the different components, whereas the deep learning based algorithm is able to effectively filter noise and artifacts and segment all the components and defects with high accuracy and at a prediction speed of about two orders faster than manual image labelling. This deep learning algorithm thus provides a promising tool to construct high-resolution numerical models for further studies such as performance optimization of C/C composites.