NYA-Crack-Data: A High Variability Concrete Crack Dataset for Enhanced Model Generalisation

Purpose:
The NYA-Crack-DATA dataset was created to address an identified gap of limited variability in existing concrete crack datasets. The crack dataset is intended to enhance the robustness of existing concrete crack datasets, thereby improving the generalisability of crack detection models.

Characteristics and variability:
The images were captured from the following locations and concrete specimens: two concrete bridges in South Wales (UK), concrete buildings around the University of South Wales (USW) Treforest Campus, indoor and outdoor concrete slabs and concrete beam, cube and cylinder specimens from laboratory experiments.

The NYA-Crack-DATA dataset was created from 297 images, captured using an iPhone 11 Pro Max, DJI Mini 3 Pro and Nikon D3400 DSLR. The images were split into smaller images (227 x 227 pixels) resulting in a total of 5026 images. The dataset consists of two classes: 'Crack' and 'No Crack', which consist of 2167 and 2859 images, respectively. The images in the dataset have not undergone any data augmentation.

The images have the following variability:
- Image quality/resolution (8000 x 6000 pixels, 6000 x 4000 pixels and 4032 x 3024 pixels)
- Lighting and weather conditions (poor and good lighting, rainy, cloudy, sunny)
- Graffiti markings
- Background noise
- Foreground occlusions such as vegetation