Is there a way to use OCR to extract specific data from a CAD technical drawing?

Question

I'm trying to use OCR to extract only the base dimensions of a CAD model, but there are other associative dimensions that I don't need (like angles, length from baseline to hole, etc). Here is an example of a technical drawing. (The numbers in red circles are the base dimensions, the rest in purple highlights are the ones to ignore.) How can I tell my program to extract only the base dimensions (the height, length, and width of a block before it goes through the CNC)?

The issue is that the drawings I get are not in a specific format, so I can't tell the OCR where the dimensions are. It has to figure out on its own contextually.

Should I train the program through machine learning by running several iterations and correcting it? If so, what methods are there? The only thing I can think of are Opencv cascade classifiers. Or are there other methods to solving this problem? Sorry for the long post. Thanks.

Why not OCR everything and build rules for what you want to extract? — Kalpit, Dec 23 '20 at 09:21
Here, I edited the question. As you can see from the picture there are so many irrelevant numbers. Plus, the numbers are not always in the same location, depending on the shape of the model, as well as the person who made the drawings. The only solution I can think of is training OpenCV to detect single arrows and identifying them as base dimensions. If there are any other solutions you can think of, it would be much appreciated. Thanks. — Joshua Park, Dec 24 '20 at 00:40

jmattes · Accepted Answer · 2022-01-14T15:19:10.063

I feel you... it's a very tricky problem, and we spent the last 3 years finding a solution for it. Forgive me for mentioning the own solution, but it will certainly solve your problem: pip install werk24


from werk24 import Hook, W24AskVariantMeasures
from werk24.models.techread import W24TechreadMessage
from werk24.utils import w24_read_sync
    
from . import get_drawing_bytes # define your own
    
    
def recv_measures(message: W24TechreadMessage) -> None:
    for cur_measure in message.payload_dict.get('measures'):
        print(cur_measure)
    
if __name__ == "__main__":
    # define what information you want to receive from the API
    # and what shall be done when the info is available.
    hooks = [Hook(ask=W24AskVariantMeasures(), function=recv_measures)]
    
    # submit the request to the Werk24 API
    w24_read_sync(get_drawing_bytes(), hooks)

In your example it will return for example the following measure

    {
        "position": <STRIPPED>
        "label": {
            "blurb": "ø30 H7 +0.0210/0",
            "quantity": 1,
            "size": {
                "blurb": "30",
                "size_type":" "DIAMETER",
                "nominal_size": "30.0",
            },
            "unit": "MILLIMETER",
            "size_tolerance": {
                "toleration_type": "FIT_SIZE_ISO",
                "blurb": "H7",
                "deviation_lower": "0.0",
                "deviation_upper": "0.0210",
                "fundamental_deviation": "H",
                "tolerance_grade": {
                    "grade":7,
                    "warnings":[]
                },
            "thread": null,
            "chamfer": null,
            "depth":null,
            "test_dimension": null,
         },
         "warnings": [],
         "confidence": 0.98810
    }

or for a GD&T

{
    "position": <STRIPPED>,
    "frame": {
        "blurb": "[⟂|0.05|A]",
        "characteristic": "⟂",
        "zone_shape": null,
        "zone_value": {
            "blurb": "0.05",
            "width_min": 0.05,
            "width_max": null,
            "extend_quantity": null,
            "extend_shape": null,
            "extend": null,
            "extend_angle": null
        },
        "zone_combinations": [],
        "zone_offset": null,
        "zone_constraint": null,
        "feature_filter": null,
        "feature_associated": null,
        "feature_derived": null,
        "reference_association": null,
        "reference_parameter": null,
        "material_condition": null,
        "state": null,
        "data": [
            {
                "blurb": "A"
            }
         ]
    }
}

Check the documentation on Werk24 for details.

score 0 · Answer 2 · answered Nov 22 '22 at 21:00

Although a managed offering, Mixpeek is one free option:

pip install mixpeek

from mixpeek import Mixpeek  
  
mix = Mixpeek(  
    api_key="my-api-key"  
)  
  
mix.upload(file_name="design_spec.dwg", file_path="s3://design_spec_1.dwg")

This /upload endpoint will extract the contents of your DWG file, then when you search for terms it will include the file_path so you can render it in your HTML.

Behind the scenes it uses the open source LibreDWG library to run a number of AutoCAD native commands such as DATAEXTRACTION.

Now you can search for a term and the relevant DWG file (in addition to the context in which it exists) will be returned:

mix.search(query="retainer", include_context=True)

[  
    {  
        "file_id": "6377c98b3c4f239f17663d79",  
        "filename": "design_spec.dwg",  
        "context": [  
            {  
                "texts": [  
                    {  
                        "type": "text",  
                        "value": "DV-34-"  
                    },  
                    {  
                        "type": "hit",  
                        "value": "RETAINER"  
                    },  
                    {  
                        "type": "text",  
                        "value": "."  
                    }  
                ]  
            }  
        ],  
        "importance": "100%",  
        "static_file_url": "s3://design_spec_1.dwg"  
    }  
]

More documentation here: https://docs.mixpeek.com/

Is there a way to use OCR to extract specific data from a CAD technical drawing?

2 Answers2