Abstract
This repository contains a dataset of images featuring concrete panels, designed to support research and development in image-based analysis and defect detection. The dataset is divided into two main categories: crack images with corresponding masks, and crack images without masks, which include both crack and no-crack images.
Technical Information
The dataset consists of concrete images originally captured at approximately 16,000 × 32,000 pixels. These images depict large concrete panels, each with dimensions of 1200 × 2000 millimeters. To facilitate the identification and extraction of specific regions containing cracks, the software ToolIP [1], especially ToolImA was utilized. The tool enabled the scanning of these large images and the selection of 224 × 224 pixel regions where cracks are present.
In total, a curated subset of 1,500 images was extracted, 1,000 of them with cracks and 500 without. Each extracted image measures 224 × 224 pixels and is presented in an 8-bit grayscale format, with the .pgm extension used for storage.
As part of the second step in the data preparation process, ToolImA was again used to perform binary pixel-wise annotations (crack pixel or background) on the selected regions. The annotations include detailed crack information, with an average crack thickness of 1–2 pixels, providing a high level of granularity for analysis and training of machine learning models. The final dataset, therefore, consists of accurately annotated 224 × 224 pixel images, formatted in 8-bit grayscale with the corresponding masks, ready for use in deep learning and other computational applications.
[1] “ToolIP – Tool for Image Processing.” Fraunhofer-Institut für Techno- und Wirtschaftsmathematik ITWM, https://www.itwm.fraunhofer.de/toolip. Accessed 20 Dec. 2024.
