Can I train a symbol as special character with ocr?

Hi there,

I'm trying to recognize a dial symbol as a special character so I can locate it when I use the ocr function.

However, is not recognizing that symbol at all.

This is my image after training the model. I wanted to locate the green symbol.

Is there another way to do this?

Do you have any suggestion @Image Analyst?

Best

Answer 1

[Shashank Gupta]

Hi Pedro,

I think you need to train an ocr to compensate for the symbol. you can refer to ocrTrain App in MATLAB. you manually have to collect enough data and train a classifier. After training you can export the file and use it in typical ocr workflow.

I hope this helps.

Cheers.

[Frank]

Hi Shashank,

I am new on matlab ocr ... I need the character "§" in base model "german" ...

I tried to extend model "german" by "§" (which is included in the "english" model, but not in german model).

I have labeled one image with two character symbols and another one with three symbols of "§", each is 25 x 37 pixels black on wthite background with good contrast.

I tried to train base model "german" (often called cdsTrain) with the 3-symbol image and "cdsValidation" with the 2-symbol image by using trainOCR-function. See gTruth-object attached ...

But the result is matlab stops (emdless loop) and must be killed ... See atached screenshot.

Do I need more training data perhaps? What went wrong with my approach?

Is there a possibility to merge "german" and "english" models to have the "§" from "english" also for german text ...?

Here is the simple code, I tried:

labelName = "Paragraph";
attributeName = "Paragr";
[imds,boxds,txtds] = ocrTrainingData(gTruth,labelName,attributeName);
cds = combine(imds,boxds,txtds);

outputDir = "C:\Data\Fuzzylogic\Mathlab\Projekte\OCR\OCR-Test\Train_Paragraph\OCRModel";
if ~exist(outputDir, "dir")
    mkdir(outputDir);
end
checkpointsDir = "C:\Data\Fuzzylogic\Mathlab\Projekte\OCR\OCR-Test\Train_Paragraph\Checkpoints";
if ~exist(checkpointsDir, "dir")
    mkdir(checkpointsDir);
end

cdsTrain = subset(cds, 1); % 3 symbols at one image 
cdsValidation = subset(cds, 2); % 2 symbols at another image

ocrOptions = ocrTrainingOptions(GradientDecayFactor=0.9, ...
    MaxEpochs=20, VerboseFrequency=160, CharacterSetSource="base-model", ...
    CheckpointPath=checkpointsDir, ValidationData=cdsValidation, ...
    OutputLocation=outputDir);

outputModelName = "paragraph";
baseModel = "german";
outputModel = trainOCR(cdsTrain, outputModelName, baseModel, ocrOptions);

Thank you very much!

Frank from Berlin