OCR (Optical Character Recognition) is a technology that systematically reads and indexes text within images. When used as part of an Engineering Document Management System such as Lunr, this can open up a large portion of the data set that would not ordinarily be retrievable using search. Searching on text-format documents like Word files or text-based PDFs (instead of image-only PDFs) loads each file into a database that supports full-text search. However, this approach only works with image files, which don't support indexing similarly.
What OCR Reads on a Drawing
A drawing is mostly linework, and its text is a small share of the sheet spread across a lot of separate places. What OCR recovers from it, in rough order of value:
- Title block text. Drawing number, sheet title, revision, discipline, scale, and the project and client names. That is the metadata the drawing already carries, so reading it means a sheet with no record behind it can still be found by its number.
- Tag and equipment numbers. The valve, pump, circuit, and room numbers labelled on the sheet itself. These are the searches that matter in operations, because somebody standing at an asset knows the tag and often nothing else.
- Notes and legends. General notes, revision notes, material callouts, and the schedule tables carrying detail the linework does not show.
- Room names and grid references. On architectural sheets, this is what makes a floor plan findable by the space it describes rather than by the building it sits in.
Two things make drawing OCR harder than OCR over a page of prose. Text sits at several rotations on the same sheet, since notes run horizontally while dimensions and pipe labels follow the geometry. And there is no surrounding sentence for a recogniser to lean on, so a tag like V-1042A has to be read character by character with nothing to check it against.
The Problem
- Historical scanned drawings (usually TIFF or PDF) often need to contain adequate metadata for search and retrieval, leading to situations where people either need to spend a lot of time browsing through archives to find the files or, worse, give up and re-draw the diagram from scratch.
- Operations and maintenance manuals, contracts, and similar files are often too large to appropriately index using standard metadata tagging techniques. Searching for the file's project name or asset description may work, but more broad searches by part description or contractor name are impossible.
Scanning the Paper Archive
Plenty of drawing archives are still paper, or microfilm, or a folder of TIFFs somebody scanned in the 1990s to a standard that made sense at the time. Scanning is the entry path to everything above, and the quality of the scan sets the ceiling on what OCR can return afterwards.
What matters when the drawings go through the scanner:
- Resolution. 300 dpi is the working minimum for a drawing. Below that, the small text in a title block and the digits in a tag number break up, and no amount of processing afterwards puts them back.
- Sheet size. Engineering drawings are A1 and A0, so this is large-format scanning. Sectioning a big sheet on an A3 scanner and joining the pieces costs more than it saves and skews the join.
- Greyscale for old sheets. A faded, stained, or pencil original scanned straight to black and white loses its light strokes. Greyscale keeps them and gives the recogniser something to work with.
- One file per sheet, named by drawing number. Scanning is the cheapest moment to attach a drawing number. Recovering it later means opening every file.
- Straight and complete. A skewed or cropped scan clips the title block, which is the one region on the sheet worth the most.
Scanning and OCR are two separate steps. The scanner produces an image, and OCR reads it. Doing the first without the second leaves a picture of a drawing, findable only by whatever somebody typed into the filename.
Reading the Archive with XRAY
Our solution is to use Lunr's XRAY technology to OCR all image files as they are uploaded. The text extracted during this process is then indexed, allowing searching over a more significant portion of your document repository. XRAY runs this way inside Lunr's drawing management software, so a scanned 1994 sheet surfaces in the same search as a live DWG.
XRAY indexes the words inside DWG, DGN, RVT, PDF, Office files, and scanned images, and returns results as rendered thumbnails so you recognise the drawing before you open it. For dense sheets, such as a schedule packed with tags or small text that a single pass under-reads, Deep XRAY re-scans the page at high resolution tile by tile and surfaces the hits the first pass missed.
Recognised text works best as the fallback behind the other ways into a repository. Tags and a curated folder structure find a drawing when somebody knows what they are looking for. Full text finds the ones nobody got around to curating, which on a legacy archive is most of them. A better way to find your engineering content sets out how the two fit together.