CrackVision12K

We present the CrackVision12k dataset, a collection of 12,000 crack images derived from 13 publicly available crack datasets. The individual datasets were too small to effectively train a deep learning model. Moreover, the masks in each dataset were annotated using different standards, so unifying the annotations was necessary. To achieve this, we applied various image processing techniques to each dataset to create masks that follow a consistent standard.Crack datasets inherently suffer from class imbalance. To mitigate this issue, we selected images containing crack pixels of more than 5000 pixels and applied data augmentation techniques such as Gaussian noise and rotation. Finally, there is a corresponding refined ground truth for each crack image across the dataset to ensure uniformity and reliability.The 13 datasets we combined are as follows: Aigle-RN, ESAR, LCMS, CRACK500, CrackLS315, CRKWH100, CrackTree260, DeepCrack, GAPS384, Masonry, Stone331, CFD, and SDNet2018.<br>Paper &amp; Code: https://github.com/junegoo94/Hybrid-SegmentorCitation:@misc{goo2024hybridsegmentorhybridapproachautomated,<br>title={Hybrid-Segmentor: A Hybrid Approach to Automated Fine-Grained Crack Segmentation in Civil Infrastructure},<br>author={June Moh Goo and Xenios Milidonis and Alessandro Artusi and Jan Boehm and Carlo Ciliberto},<br>year={2024},<br>eprint={2409.02866},<br>archivePrefix={arXiv},<br>primaryClass={cs.CV},<br>url={https://arxiv.org/abs/2409.02866},<br>}<pre></pre><br>

本研究提出CrackVision12k数据集，该数据集包含12000张裂缝图像，源自13个公开可用的裂缝数据集。单个数据集规模过小，无法有效支撑深度学习模型的训练。此外，各数据集内的裂缝掩码（mask）采用了不同的标注标准，因此亟需统一标注规范。为此，我们针对每个数据集应用多种图像处理技术，生成符合统一标准的掩码。裂缝数据集本身存在类别不平衡问题。为缓解该问题，我们筛选出包含5000个以上裂缝像素的图像，并应用高斯噪声、旋转等数据增强技术。最终，数据集中的每张裂缝图像均配有经过精修的真实标注（ground truth），以保障标注的统一性与可靠性。本次整合的13个数据集如下：Aigle-RN、ESAR、LCMS、CRACK500、CrackLS315、CRKWH100、CrackTree260、DeepCrack、GAPS384、Masonry、Stone331、CFD以及SDNet2018。
论文与代码：https://github.com/junegoo94/Hybrid-Segmentor
引用：@misc{goo2024hybridsegmentorhybridapproachautomated,
title={Hybrid-Segmentor: A Hybrid Approach to Automated Fine-Grained Crack Segmentation in Civil Infrastructure},
author={June Moh Goo and Xenios Milidonis and Alessandro Artusi and Jan Boehm and Carlo Ciliberto},
year={2024},
eprint={2409.02866},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2409.02866},
}

搜集汇总

背景与挑战

背景概述

CrackVision12K是一个包含12,000张裂缝图像的数据集，整合了13个公开数据集，通过统一标注标准和数据增强技术解决了原始数据集的小规模、标注不一致和类别不平衡问题。