LB Inpainting Dataset

Images inpainted by Lattice Boltzmann based anisotropic diffusion method.
This dataset was created to support research in the fields of image inpainting, defect detection, and restoration algorithms. It consists of 10,000 grayscale images originally selected from a larger public dataset Places* and converted to grayscale. Synthetic defects were introduced into these images to simulate realistic degradation, and corresponding binary masks were generated to annotate the defect regions. Additionally, the dataset includes the results of applying an image inpainting algorithm to these corrupted images, allowing researchers to evaluate restoration quality.

The dataset is structured into four main folders:

"orig.zip": Contains the original 10,000 grayscale images. These were converted from color images selected from a public dataset, resized and preprocessed uniformly.

"spot.zip": Contains the same images as `orig/`, but with synthetic defects applied. The defects are designed to imitate common issues in industrial inspection, such as occlusions, stains, or noise artifacts.

"mask.zip": Contains binary masks corresponding to the defect locations in the `spot/` images. A value of `1` indicates a defect pixel, and `0` indicates background.

"result.zip": Contains the inpainted images generated from the `spot/` images using a selected inpainting algorithm. These outputs can be used to evaluate inpainting performance using both visual inspection and quantitative metrics.

*Places: A 10 million Image Database for Scene Recognition
B. Zhou, A. Lapedriza, A. Khosla, A. Oliva, and A. Torralba
IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017