How do I import a table containing numbers in a picture with OCR?

I have a picture of a table with numeric data in it. Like this:

How do I import the values with ocr? Is there a better method?

I tried to use ocr, but it's not working properly.

Thank you

[Mohammad Sami]

If images is all you have, ocr would be the easy way.

try to restrict the characters to between 0 to 9 and - , .

[Mattia Tosi]

I will try, but sometimes it's a mix of characters and numbers. How can I capture both? It looks like ocr is not really effective...

[Mohammad Sami]

Another option, that might improve the ocr accuracy sometimes is to imresize the input image.

Answer 1

[Mohammad Sami]

Try resizing the image. It would hopefully improve the accuracy.

a = imread('image.jpeg');
a = imresize(a,2);
txt = ocr(a,'CharacterSet','0123456789.');
Iocr = insertObjectAnnotation(a, 'rectangle', ...
txt.WordBoundingBoxes, ...
txt.WordConfidences);
for i = 1:length(txt.Words)
Iocr = insertText(Iocr,txt.WordBoundingBoxes(i,1:2),txt.Words{i},'AnchorPoint','Center');
end
figure; imshow(Iocr);

[Mattia Tosi]

Limiting the 'CharacterSet' to '0123456789.' helps but imresize in my case does not give the expected results. (I can't disclose the table I am working on unfortunately).

However, this partially solves my problem. In fact I still need to capture alphabetic characters, but I need to improve the detection of numeric elements.

[Mohammad Sami]

If you are really having a lot of issue, you can train your own ocr model using the included ocr trainer app.

However you will need to manually label the training set in the ocr trainer app. The app would label the data initially, you can review and correct the mistakes. Then train the model using the newly labelled data.

You can also try the various matlab image enhancements, to see if any of those help.

[Mohammad Sami]

If you know the font that was used to produce your images. you can follow this tutorial

https://www.mathworks.com/help/vision/ug/train-optical-character-recognition-for-custom-fonts.html

[Mattia Tosi]

It doesn't work direcly with the table I have, but playing with all the parameters you mentioned the result is improving.