Structural Defects Network — Images of cracks in concrete
The Structural Defects Network dataset (SDNET2018) contains annotated images of cracks in concrete surfaces (bridges, walls, roads). It includes very fine to large cracks, with various lighting conditions and obstacles, to train detection models based on deep learning.
56,000+ images segmented into 256×256 px sub-images, JPEG/PNG format
License not explicitly specified, dataset freely accessible
Description
The dataset SDNET2018 provides more than 56,000 annotated images of cracks on concrete surfaces, captured in various contexts (bridges, walls, sidewalks). Each image is segmented into 256x256 pixel sub-images that are labeled “cracked” or “uncracked.” The dataset is ideal for training algorithms for the automatic detection of cracks using convolutional neural networks.
What is this dataset for?
- Develop automated systems for the visual inspection of civil infrastructures
- Training deep learning models for the early detection of structural defects
- Evaluate and compare image analysis techniques for structural health monitoring
Can it be enriched or improved?
Yes, the dataset can be enriched by additional precise annotations (e.g. types of cracks, depth), or by increasing the images (lighting variations, angles). An extension to other materials would also be beneficial.
🔎 In summary
🧠 Recommended for
- Industrial AI researchers
- Civil engineers
- Computer vision students
🔧 Compatible tools
- TensorFlow
- PyTorch
- OpenCV
- LabelImg
💡 Tip
Use image augmentation techniques to improve robustness to changes in lighting and angles.
Frequently Asked Questions
What is the resolution of the images in the dataset?
The images are segmented into square sub-images of 256x256 pixels for easy annotation and training.
Does this dataset include different types of concrete surfaces?
Yes, it contains images of bridges, walls, and sidewalks with various textures and environmental conditions.
Can this dataset be used to train a detection model in real time?
Yes, images and annotations are suitable for developing efficient models in real time with lightweight CNN architectures.
