P→MPDFMDSystem documentation
Format referenceOpen converter ↗

Level 4 · Recognition method

Last modified: (dates/navigation repair). Content current as of: (last retained content commit; content verification date unavailable). Original content last edited: . Current claims have not been freshly reverified.

Tables, graphics, and unknown regions

Tables are recognized as relationships among cells, not simply as aligned text. Meaningful graphics remain separate, addressable evidence rather than being flattened into prose.

On this page

Table model

One cell-adjacency graph

Visible rules

Vector or raster lines, intersections, rectangles, fill regions, and text containment. Multi-scale scan analysis can join broken rules without treating ordinary underlines as table borders.

Borderless structure

Repeated alignment stops, shared row baselines, whitespace gutters, column value types, compact gaps, header contrast, and recurring row patterns.

Long flowing lines, weak row synchrony, paragraph-shaped gaps, and unstable column roles support multi-column prose instead. Nearby text joins a table only when it improves the row graph; proximity alone is insufficient.

Simple rectangular content uses a GFM table. Spanning cells, nested content, shading, and page fragments use restricted HTML inside a PDFMD table directive. Empty cells are explicit when supported by the grid.

Across pages

Table-fragment fingerprints

Adjacent fragments are compared by normalized column edges, column count, value types, header text and style, ruling or fill style, width, caption relation, and boundary position. Repeated headers become header occurrences, not data rows. Totals, a terminal caption, a new heading, or incompatible columns block a join.

Ambiguity handling. If geometry and row semantics do not justify a join, fragments remain separate with a possible-continuation relation.

Meaningful visual evidence

Signatures, handwriting, stamps, and seals

ClassSupporting observations
SignaturePen-like stroke topology, aspect and ink behavior, nearby signature lines or signatory fields, and signing context.
Stamp or sealClosed-frame geometry, radial or rectangular structure, OCR inside the frame, overprint, and ink.
HandwritingIrregular strokes, handwriting proposals, color clusters, and local form context.

Each object keeps a stable ID, crop or asset reference, page, bounding geometry, class confidence, measured ink, optional OCR text, and relations. Associating a signature with a person is a separate decision; only explicit source evidence or human confirmation makes it confirmed.

Coverage

Unrecognized does not mean discardable

Raster regions, unexplained paint, and graphical candidates remain represented when their semantic class is unresolved. Every meaningful source region must be represented, suppressed with a recorded reason, or marked unknown or failed.