OCR: How to extend the character set for german language OCR model set by additional characters?

My documents in german language contain "§". The character set of the model of german language pack used for OCR does not contain this character, but english model does ... The aim is to enable OCR to recognize this special character in german language documents.

How to extend the character set of german laguage model by "§" (or another character)?

I checked a lot of search requests but did not found an example or an answer to solve this.

Thanks a lot!

Frank

Answer 1

[Sreeram]

Hi Frank,

To enable OCR to recognise the "§" character in German language documents, you may train a new OCR model on "§" by fine-tuning the German model using the “trainOCR” function. Detailed instructions can be found in the following documentation:

https://www.mathworks.com/help/releases/R2024b/vision/ref/trainocr.html?#d126e301790

If training a new model is not feasible, you may specify multiple models (German and English) to use for detection simultaneously. This can be done by passing them as a cell array in “Model” argument of the “ocr” function as specified in the following documentation:

https://www.mathworks.com/help/releases/R2024b/vision/ref/ocr.html#bu76sfz

ocr(I,model={"german","english"})

Note that this approach might result in some German characters being misclassified as English.

I hope this helps!

[Frank]

Hi Sreeram,

thank you very much for your answer. But both suggestions do not work:

1) If I train the german model by labeling many "§"-chars, exporting the gTruth data with set labels attribute values to "§", matlab hangs and must be "killed" (may be it cannot handle the not known character "§" ...).

On occasion: What's the meaning of all the "joined"-Phrases in OCR text results when using self trained data?

2) The proposal of model={"german","english"} delivers just the same results like using only the english model ... (no differences).

Do you have another idea?

If not, I will do two separate OCRs (the german and the english), will identify the positions of all "§" in english text and replace it within the german OCR-text. It's a little bit laborious but probably will be a working workaround.

I want to use Matlab OCR for validating german juridical documents and there are a lot of "§" signs within ... To navigate trough the documents texts by using different fuzzy inference systems to find out the semantics I need to know where "§"s are located.

Thanks a lot for the great service! And by the way - Matlab is really great. I do programming for more than 30 years, so I know.

Frank from Berlin

[Frank]

Hi Sreeram and others interested in ..

I copied the file "deu.traineddata" from folder "C:\ProgramData\MATLAB\SupportPackages\R2024b\3P.instrset\tesseract-ocr-languages-deu.instrset\tessdata_best" to folder "D:\Programs\MATLAB\R2024b\toolbox\vision\visionutilities\tessdata_best" and now it works (line: "result = ocr(I,LayoutAnalysis="page",Model={"german","english"})")

In my windwos 11 Matlab installation the program folder is on disk D:\... but the support packages were automatically installed to C:\... May be this is not the expected standard.

Thanks a lot!

[Frank]

Important notice for all interested in: The sequence of model names in the statement is important.

Model={"english","german"} works much more better than Model={"german","english"} ...

Final message now ;-)