CSEM-MISD - CSEM's Multi-Illumination Surface Defect Detection Dataset

In automated surface visual inspection, it is often necessary to capture the inspected part under many different illumination conditions to capture all the defects. To address this issue, at CSEM we have acquired a real-world multi-illumination defect segmentation dataset, called CSEM-MISD and we release it for research purposes to benefit the community.

The dataset consists of three different types of metallic parts -- washers, screws, and gears (temporarily only the first one is available). Parts were captured in a half-spherical light-dome system that filtered out all the ambient light and successively illuminated it from 108 distinct illumination angles.  Each 12 illumination angles share the same elevation level and the relative azimuthal difference between the adjacent illumination angles on the same level is 30 degrees. For more details, please read Sections 3 and 4 of our paper.

The washers dataset features 70 defective parts. Some defects, such as notches and holes, are visible in most images (illuminations) with intensity and texture variations among them, while others, such as scratches, are only visible in a few.

We split the datasets into train and test sets. The train sets contain 32 samples, and the test set 38 samples. Each sample comprises 108 images (each captured under a different illumination angle), an automatically extracted foreground segmentation mask, and a hand-labeled defect segmentation mask.

This dataset is challenging mainly because:

• each raw sample consists of 108 gray-scale images of resolution 512×512 and therefore takes 27MB of space;
• the metallic surfaces produce many specular reflections that sometimes saturate the camera sensors;
• the annotations are not very precise because the exact extent of defect contours is always subjective;
• the defects are very sparse also in the spatial dimensions: they cover only about 1.4% of the total image area in washers; this creates an unbalanced dataset with a highly skewed class representation. 

The dataset is organized as follows:

• each sample resides in the Test, Train, or Unannotated directory;
• each sample has its own directory which contains the individual images, the foreground, and defect segmentation masks;
• each image is stored in 8-bit greyscale png format and has a resolution of 512 x 512 pixels;
• Image file names are formatted using three string fields separated with the underscore character: prefix_sampleNr_illuminationNr.png, where the prefix is e.g. washer, the sampleNr might be a three-digit number 001, and the illuminationNr is formed of 3 digits, first corresponding to the elevation index (1 - highest angle, 9 - lowest angle), and the additional two corresponding to the azimuth index (01-12).
• Each dataset contains light_vectors.csv, which contains the illumination angles (in lexicographic order of the illuminationNr), and light_currents.csv that contains the numbers roughly corresponding to the light intensity.

We provide data loaders implemented in python at the project's repository.

If you find our dataset useful, please cite our paper:

Honzátko, D., Türetken, E., Bigdeli, S. A., Dunbar, L. A., & Fua, P. (2021). Defect segmentation for multi-illumination quality control systems. Machine vision and Applications.