Extracting Title Block Data From an Engineering Drawing

A title block is the one part of an engineering drawing whose location you can rely on. Its contents are another matter: two drawings of the same part, one drawn to ISO conventions and one to ASME, will agree on about half their fields and disagree about a symbol that determines whether the views are mirrored.

Where it is, and why that helps

Sheet layout is standardised — ISO 5457 for ISO practice, ASME Y14.1 for inch drawing sheets — and both put the title block in the lower right corner of the sheet, reading in the same orientation as the drawing. That is a stronger constraint than it sounds. It means you can crop the region before you do anything else, and cropping is the single most effective intervention available on a drawing, because the rest of the sheet is a dense field of geometry, dimensions and notes that nothing in the pipeline wants to read.

Crop to roughly the lower right quarter, at full resolution rather than a downscaled whole sheet. A drawing rendered to fit a model’s image size limit puts the title block text below legibility — this is the same resolution trade-off described on the page about image detail levels, and on E-size and A0 sheets it is decisive rather than marginal.

The corner also holds things that are not the title block: a revision block, a parts list or bill of materials, general notes, and an application block showing the next assembly. Cropping to the corner gets you all of them, which is convenient if you know they are separate tables and confusing if you ask for “the title block” and get a parts list merged into it.

The ISO 7200 field set

ISO 7200:2004 specifies data fields for title blocks and document headers, and it is unusually concrete: it defines eight mandatory data fields and eleven optional ones. The mandatory set is legal owner, identification number, date of issue, segment or sheet number, title, approval person, creator, and document type.

Two of those are worth pausing on. The identification number is required to be unique at least within the legal owner’s organisation, which makes it the key for the extracted record — not the title, which is descriptive and frequently duplicated across a family of parts. And legal owner is often present only as a logotype, because the standard permits it: a company mark in the corner of the block is a populated mandatory field, not a missing one, and a text-only extraction will report it as absent.

The optional eleven cover the fields people actually argue about — revision index, sheet count, language code, responsible department, technical reference, document status, classification. Optional means a conforming drawing may omit them, so a schema that marks scale or material as required will fail on drawings that are entirely correct — the case for designing a schema for variants you have not seen.

What an ASME drawing does differently

ASME practice is spread across several documents rather than one: Y14.1 for sheet size and format, Y14.100 for engineering drawing practices generally, Y14.35 for revisions. The observable differences in the block itself are the ones to encode.

• Units. An ASME drawing commonly states units in a note near the block rather than as a labelled field, and the default may be inches with no statement at all. An ISO drawing is millimetres unless it says otherwise. A dimension of 2.500 means two entirely different sizes depending on which you are holding.
• General tolerances. ISO drawings usually cite a general tolerance standard, most often ISO 2768 with a class letter such as -m or -f. Inch drawings usually print a decimal-place tolerance block instead — one tolerance for .X, another for .XX, another for .XXX. These are the same information in incompatible forms and neither converts to the other automatically.
• Third-party approvals. Drawn by, checked by, approved by, plus quality and engineering signatures, each with a date. ISO 7200 makes one approval person mandatory; drawings in practice carry several, so extract them as a list of role, name and date.

The field with no text

The projection-angle symbol is a small pictogram in or beside the title block: a truncated cone drawn with a circle and a trapezoid, arranged one way for first-angle projection and mirrored for third-angle. First angle is the ISO convention and third angle is the ASME one. It has no text, so no text extraction returns it, and no schema that lists “fields” will ask for it.

It also matters more than most of the fields that do have text. The projection convention determines which side of the front view the right view is drawn on. Read a first-angle drawing as third-angle and the part is effectively mirrored — a mistake that survives right up until somebody machines it. If your extraction feeds anything that reasons about geometry rather than just filing the document, this needs a vision pass over the cropped symbol region with an explicit two-value answer, and an unknown that is allowed to be returned rather than guessed, which is how an illegible field should be handled generally.

The same is true of a handful of other glyph-only markings: the surface texture symbol, the weld symbol conventions, and the “do not scale drawing” note that is sometimes printed as a struck-through scale field rather than as words.

Why the text comes out scrambled

• The block is drawn, not typeset. In a CAD-exported PDF the title block is line geometry with independently positioned text objects. There is no table structure, no cell boundaries and no reading order — the text objects come out in whatever order the exporter emitted them, which is often creation order. Labels and values must be paired by coordinate proximity, not by sequence.
• Label and value are separate objects. “SCALE” and “1:2” are two unrelated strings that happen to be near each other. A model given the raw text stream sees a bag of words; a model given the cropped image sees a table. For this specific block, the image route is usually the better one even when a text layer exists.
• Rotated text. Sheet borders, zone markings and occasionally block fields are rotated ninety degrees. Extractors differ on whether they emit rotated text at all.
• Scale is not always a ratio. 1:2, 2:1, NTS, NONE, AS NOTED and a blank are all values that appear. Model it as a string with an optional parsed ratio, not as a number.
• Sheet 1 of 4 is a relationship. The sheet number and count tie the drawing to its siblings. Extracted per file with no link between them, a four-sheet drawing becomes four documents that each claim to be the whole thing — and a missing sheet becomes invisible, which is the problem the revision history page attacks from the other end.