Extracting feature information from 2D CAD engineering drawings

edit

The use-case I have is, that thousands of such drawings exist, mostly as PDF, some even only as raster image and there is no chance to access any original files.

1. is no problem, because I focus on legacy data. For new stuff, machine-readable data is simpler to get.

2. 2.-4. are all valid points and need to be handled.

Respectfully (not to deter you)... with only thousands and so many criteria, would you not use a temp/intern? Data would need validation regardless.

Yes, that's what we currently propose. But companies love “let's automate this" and there can be around 500.000 drawings. If all these need to be analyzed (automatically or manually) is a different story.

Make a portal. List the documents with status (new,pending,approval,error,done). Have a sidebar with sets of visual reference symbols/tools. Develop software as you make progress for real feedback and results. Farm the annotation/review out. Maybe there's a onionskin with original, OpenCV detection, and a cross-reference to editable database fields. Get production startup semi-automatic, always progress...

Also, does this service exist at $0.01/document, or should you protect any future copyright/development? For example, upsample documents, is that in scope?

Sorry for being that late...

We had the portal idea too. Would be a desktop tool where one captures some data directly...

Don't understand your last questions.

The idea is that companies can use this tool in-house and build-up a shared-service or use a shared-service provides who use the tool.

Yes, I was curious if your intent was SaaS. Advancing from a question of high-level feature requirements to implementation... i.e. a Java developer with a background in drafting.

• whitelist symbol library / dictionary
• blacklist patterns, i.e. hatching
• internationalization
• preflight routine i.e. upscale img or convert pdf
• interactive annotation / hints for interpreter to refine output
• multiple view / detail consolidation
• text(s) recognition
• unit conversion

That's a good rough summary :-)

Are you aware of any OpenCV examples, that analyze such CAD drawings?

I'm not sure yet, how much AI training approach is needed, or if such information can be handled with an algorithmic approach or if a mixture is a way to go. And for an AI approach, how many training data will be necessary. Any experience?

There are many text recognition examples, classifiers, etc. in multiple languages. Maybe thousands of labelled training data to start? I am only learning myself, but I have been in startups... "window-shopping" solutions. The pace of progress has people turning to cloud services and certified specialists, but you still need the language to specify. Granted, there is fluidity of expectation and budget, staff. So maybe it's a Mechanical Turk at first and then it evolves. A viable POC.

Do any modern CAD programs offer or project this service? Adobe? It seems within scope.