Cracks in Steel Bridges (CSB) dataset: data underlying the publication: Loss function inversion for improved crack segmentation in steel bridges using a CNN framework

The presented dataset used for the experiments is described in the article "Loss function inversion for improved crack segmentation in steel bridges using a CNN framework" (doi:https://doi.org/10.1016/j.autcon.2024.105896). The dataset consists of images of steel bridge structures and pixel-wise fatigue crack annotations. Some of the images contain bridge structures with cracks or corrosion, while others capture structures without any defect. The images are provided by bridge infrastructure owners "Rijkswatersaat" and "ProRail" and by "Nebest" engineering company. The annotation of images was made using a semi-automatic annotation tool described in the article "Segmentation Tool for Images of Cracks" (doi:https://doi.org/10.1007/978-3-031-35399-4_8) and which implementation is available at https://github.com/akomp22/crack-segmentation-tool.The dataset consists of high-resolution images and is stored in the folder "entire images". The images are divided into test and train sets. Images that capture cracks are stored in the folder "crack_train" and "crack_test". Images capturing structure without a crack are stored in folders "nocrack_train" and "nocrack_test". For each image, a .json file is stored in the same folder and under the same name as the corresponding image. The .json file stores the position (x,y) of pixels on the image, which lie in a crack region. An example of a code to generate a binary segmentation map from the .json files is given in the "read_json_annotation.py" file.Additional patch datasets were generated from the entire images. The patch datasets are stored in the “patch dataset” folder. The multiple patch datasets differ by the patch size, number of patches, and fraction of patches that do not contain cracks among all patches of the particular dataset. Furthermore, we provide segmentation maps in file "predictions.rar" for entire test images which are given by the method proposed in our research article.For more explanations, please refer to the article: https://doi.org/10.1016/j.autcon.2024.105896

本实验所用数据集已在论文《基于卷积神经网络（Convolutional Neural Network, CNN）框架的损失函数反演以优化钢结构桥梁裂纹分割》（DOI：https://doi.org/10.1016/j.autcon.2024.105896）中进行了详细阐述。该数据集包含钢结构桥梁图像以及逐像素疲劳裂纹标注信息。部分图像拍摄了带有裂纹或腐蚀病害的桥梁结构，其余图像则记录了无任何缺陷的桥梁结构。
图像由桥梁基础设施业主Rijkswatersaat、ProRail，以及Nebest工程公司提供。图像标注采用了论文《图像裂纹分割工具》（DOI：https://doi.org/10.1007/978-3-031-35399-4_8）中介绍的半自动标注工具，该工具的开源实现地址为https://github.com/akomp22/crack-segmentation-tool。
该数据集包含高分辨率图像，存储于“entire images”文件夹中。图像被划分为训练集与测试集：带有裂纹的图像分别存储于“crack_train”与“crack_test”文件夹内，无裂纹的结构图像则存储于“nocrack_train”与“nocrack_test”文件夹中。
针对每张图像，对应文件夹下会存在一个与图像同名的.json文件，该文件存储了图像中属于裂纹区域的像素的坐标(x,y)。在“read_json_annotation.py”文件中，提供了一段用于从.json文件生成二值分割图的示例代码。
此外，研究团队还从完整图像中生成了多组补丁数据集，此类数据集存储于“patch dataset”文件夹中。不同补丁数据集的差异体现在补丁尺寸、补丁数量，以及特定数据集全部补丁中无裂纹补丁的占比。
同时，本研究提出的方法生成的全测试集图像分割结果已打包存储于“predictions.rar”压缩包内。如需更多详细说明，请参阅该论文：https://doi.org/10.1016/j.autcon.2024.105896