The authors create the Fiber Segmentation Dataset, a small dataset to segment fibers in CT scans of concrete. The created fibers dataset consists of only 3 spatially disjoint volumes of size 20 x 512 x 512 (d x h x w) voxels (voxel size: 4 µm). It was geometrically enlarged by combinations of rotation (using multiple angles), resizing, flipping, tilting, and squeezing using the AiSeg project.

Motivation

To optimize material usage in construction projects, carbon reinforced concrete presents a viable solution. However, the strategic placement of carbon elements becomes pivotal in this approach. To address this, the researchers employed a sophisticated micro-tomography instrument known as µ-CT. This apparatus comprises an X-ray source and a camera, which capture numerous projections while the object undergoes rotation. Subsequently, these projections are utilized to generate a comprehensive volumetric reconstruction of the object. Within these volumetric representations, the primary task involves isolating the carbon components, a task efficiently executed through convolutional neural networks (CNNs). Typically, volumetric datasets consist of a multitude of 2D images stacked sequentially, where each pixel in a 2D image corresponds to a voxel in three-dimensional space. These datasets were gathered using various acquisition devices including magnetic resonance imaging, electron microscopy, and computed tomography, each presenting unique challenges that the neural networks must overcome. Furthermore, to expand the dataset repertoire, new datasets encompassing carbon rovings, concrete pores, and polyethylene fibers were generated using Dragonfly software.

Dataset description

Segmenting polyethylene fibers within strain-hardened cement-based composites poses a considerable challenge due to the striking similarity in density between individual fibers of carbon or polyethylene and quartz sand particles, resulting in nearly identical gray values. The dataset for polyethylene (PE) fibers comprises merely three spatially disjoint volumes, each measuring 20 x 512 x 512 voxels (d x h x w) with a voxel size of 4 µm. Given the small scale of this dataset, geometric augmentation techniques were employed to enhance its diversity, including rotations at multiple angles, resizing, flipping, tilting, and squeezing using the AiSeg project. This augmentation was crucial to prevent overfitting, considering the limited number of original images available for training. A total of 397 training and 100 validation volumes were generated, while the original volumes remained designated for testing purposes. Notably, the geometric augmentation resulted in varied shapes across the newly created volumes. This dataset presents a unique challenge of accurately predicting real volumes using minimal and augmented data, further compounded by the inclusion of predominantly thin objects within the dataset.

Miniature without labels and labeled volume containing fibers (blue).