A step-by-step C# implementation of the Docstrum algorithm for pdf documents
https://devblogs.microsoft.com/cesardelatorre/using-ml-net-in-jupyter-notebooks/
https://devblogs.microsoft.com/dotnet/net-interactive-is-here-net-notebooks-preview-2/ (if the previous fails when installing)
Version 2 is the latest update and handles rotated words/lines/paragraphs. This is the version implemenented in PdfPig.
Link to original paper: The Document Spectrum for Page Layout Analysis by Lawrence O'Gorman
From Performance Comparison of Six Algorithms for Page Segmentation: The Docstrum algorithm by O'Gorman is a bottom-up approach based on nearest-neighborhood clustering of connected components extracted from the document image. After noise removal, the connected components are separated into two groups, one with dominant characters and another one with characters in titles and section heading, using a character size ratio factor fd. Then, K nearest neighbors are found for each connected component. Then, text-lines are found by computing the transitive closure on within-line nearest neighbor pairings using a threshold ft. Finally, text-lines are merged to form text blocks using a parallel distance threshold fpa and a perpendicular distance threshold fpe. - wiki
The variables can be accessed by using the GetStructuralBlockingParameters() function.
public static bool GetStructuralBlockingParameters(PdfLine i, PdfLine j, double epsilon,
out double angularDifference, out double normalisedOverlap, out double perpendicularDistance)
{
...
}From the original paper by O'Gorman:
Fig. 8. Variables used in structural block determination. The two text lines, represented by segments i and j, are to be tested here to determine if they should be grouped into the same block. Their angular difference is θij. The overlap length of segment i on segment j is pj, (and that is normalized to obtain the overlap parameter). The parallel distance between i and j is daij = pj in this case. The perpendicular distance betwen i and j is deij.
