Vickers Indentation Database and Ground Truths

Paper

Vickers Indentation Database and the Ground Truths

On this page we provide:

• Vickers Indentation Database
• Manual (Ground Truth) Diagonal length Measurements
• Manual (Ground Truth) Indentation Area Masks

Vickers Indentation Database is composed of two data sets in which 150 images are included in DB-I and 216 images in DB-II. All images have been taken from camera equipped hardness testing machines in industrial production environments. Each machine has an optical microscope that offers ten- to hundredfold magnification of the specimen and a monochrome CCD camera that produces 8-bit grey scale images at a resolution of 1280 x 1024 pixels. Although originally in JPEG format, the images have been converted to PNG format during collection from the system. A flaw of the camera system is that the border pixels of the image can be damaged an shall thus be ignored. Indentation images for the database are selected to provide a wide range of variety and to include a reasonable number of challenging surfaces, however they always show exactly one Vickers indentation. The images have a significant variation in size, position and rotation of the indentation, the texture of the specimen and the contrast of the picture.

Manual (Ground Truth) Diagonal length Measurement is carried out by an operator attaching measurement lines to the indentation images. If necessary, theses lines are rotated to mimic a rotation of the indentation in the image. The distances of corresponding measurement lines give the diagonals, which are further used for the calculation of the hardness value. To achieve stable results, the measurements have been carried out not only by one person but by a number of hardness testing experts. DB-I has been measured by four and DB-II by five different people. These independent measurements show surprisingly high variations in the position and the rotation of the measurement lines and thus the lengths of the diagonals. Therefore, for diagonal lengths, to level out human measurement errors, each diagonal length produced by algorithms should be compared to the median of diagonal lengths of corresponding manual measurements as the error metric. A similar approach is to be taken when rating vertex positions of the vertices produced by the automated indentation

Manual (Ground Truth) Indentation Area Mask is used to train superwised learning models (e.g. CNN). We generated two sets of ground truth data for both datasets. For this purpose we used an image labeling tool to articulate the indentation pixels using two approaches: In the first approach, the indentation pixels are marked by fitting a rectangle-shaped mask to the indentation vertexes. While this approach provides a good approximation of the indentation area, yet, due to the fact that the actual indentations' borders have rather concave or convex profile, in some cases the fitted mask on the indentation areas includes some background (surface) pixels or vise versa. To address this, in the second approach we attach polygon vertices to the indentation borders points, generating more precise groundtruth masks. As a further contribution of this paper, both ground truth sets are made publicly available2 as well.

If you are using the above database or the ground truths in your work, please cite the paper:

Bibtex

@Article{ Jalilian24a,
      author = {Ehsaneddin Jalilian and Andreas Uhl},
      title = {A Benchmark for Automated Vickers Hardness Testing", In Fifth Workshop on Reproducible Research in Pattern Recognition (satellite event of ICPR 2024)},
      journal = {In Fifth Workshop on Reproducible Research in Pattern Recognition (satellite event of ICPR 2024)},
      year = {2024},
      publisher = {Springer},
      pages = {1-8}
}