Numbers detection

[Aftab]

Hey guys,
I am very new to image processing & OCR. But after a lot of trial and error.
I have reached to this point. I have a small image, cropped from larger input and the image is pre-processed to maximise the visibility of the number.

It is able to detect 10000 at the top, but it is not able to detect the number on the bottom.
Here is the processed image I am working with.

.

I am running this in Browser using the tesseract.js node module, and here is my code for the detection: Tried with default pageseg_mode, as well as various other modes. 11 worked best out of all.

async function recognizeText(image) {

const worker = await createWorker('eng');

await worker.setParameters({

tessedit_char_whitelist: '0123456789',

tessedit_pageseg_mode: '11',

});

const ret = await worker.recognize(image);

console.log(ret.data.text);

await worker.terminate();

}

[Zdenko Podobny]

Hello,

For Tesseract you need to remove all non-text parts (graphics element). IMO also the outline number would be problematic. 

It would be better to post the original image so people can play with preprocessing...

See e.g. this discussion https://groups.google.com/g/tesseract-ocr/c/YqW9XhbWC_8/m/75juLKoJDwAJ (not sure if this is possible with javascript)

Zdenko

ut 19. 12. 2023 o 17:08 Aftab <afta...@gmail.com> napísal(a):

--
You received this message because you are subscribed to the Google Groups "tesseract-ocr" group.
To unsubscribe from this group and stop receiving emails from it, send an email to tesseract-oc...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/7e3e4683-2c35-42d2-ba72-8df2773d15b9n%40googlegroups.com.

[Pankaj Duggal]

Hello 

Can you help us with what steps needs to be taken with image

Fonts size are unequal 

Somewhere background colors are dark text are light 

What can we increase the accuracy in this ?

[Aftab]

Thank you for the response.
This is my original image, any pointers on how to remove the non-text part would be helpful.

[Art Rhyno]

If your environment supports something like opencv, and if the numbers are a consistent color, you could try to get the bottom numbers by leveraging the color to extract them. I have attached a simple example from using the code below.

 

Best,

 

art

---

import cv2

import numpy as np

image = cv2.imread('test.png')

 

lower_white = np.array([200, 200, 200], dtype = "uint8")

upper_white= np.array([255, 255, 255], dtype = "uint8")

 

mask = cv2.inRange(image, lower_white, upper_white)

detected_output = cv2.bitwise_and(image, image, mask = mask)

cv2.imwrite("white.png",detected_output)

 

From: tesser...@googlegroups.com <tesser...@googlegroups.com> On Behalf Of Aftab
Sent: Wednesday, December 20, 2023 7:01 AM
To: tesseract-ocr <tesser...@googlegroups.com>
Subject: Re: [tesseract-ocr] Numbers detection

 

You don't often get email from afta...@gmail.com. Learn why this is important

Thank you for the response.
This is my original image, any pointers on how to remove the non-text part would be helpful.

On Wednesday, December 20, 2023 at 10:56:16 AM UTC+5:30 pankaj....@gmail.com wrote:

Hello 

 

Can you help us with what steps needs to be taken with image

 

Fonts size are unequal 

Somewhere background colors are dark text are light 

 

What can we increase the accuracy in this ?

 

On Tuesday 19 December 2023 at 21:45:18 UTC+5:30 Zdenko Podobny wrote:

Hello,

 

For Tesseract you need to remove all non-text parts (graphics element). IMO also the outline number would be problematic. 

 

It would be better to post the original image so people can play with preprocessing...

See e.g. this discussion https://groups.google.com/g/tesseract-ocr/c/YqW9XhbWC_8/m/75juLKoJDwAJ (not sure if this is possible with javascript)

Zdenko

 

 

ut 19. 12. 2023 o 17:08 Aftab <afta...@gmail.com> napísal(a):

Hey guys,
I am very new to image processing & OCR. But after a lot of trial and error.
I have reached to this point. I have a small image, cropped from larger input and the image is pre-processed to maximise the visibility of the number.

 

It is able to detect 10000 at the top, but it is not able to detect the number on the bottom.
Here is the processed image I am working with.

.

I am running this in Browser using the tesseract.js node module, and here is my code for the detection: Tried with default pageseg_mode, as well as various other modes. 11 worked best out of all.

async function recognizeText(image) {

const worker = await createWorker('eng');

await worker.setParameters({

tessedit_char_whitelist: '0123456789',

tessedit_pageseg_mode: '11',

});

const ret = await worker.recognize(image);

console.log(ret.data.text);

await worker.terminate();

}

--
You received this message because you are subscribed to the Google Groups "tesseract-ocr" group.
To unsubscribe from this group and stop receiving emails from it, send an email to tesseract-oc...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/7e3e4683-2c35-42d2-ba72-8df2773d15b9n%40googlegroups.com.

--
You received this message because you are subscribed to the Google Groups "tesseract-ocr" group.
To unsubscribe from this group and stop receiving emails from it, send an email to tesseract-oc...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/417e7c17-0ef4-4cb2-a7e9-2ae707c91cbbn%40googlegroups.com.

[Ger Hobbelt]

On Thu, 21 Dec 2023, 15:22 Art Rhyno, <artr...@uwindsor.ca> wrote:

If 

Important extra note (as I see a new image that's white text on black background):

Tesseract was trained on black text on white background, targeting books, publications and academic papers' OCRing.

To improve your chances, ALWAYS make sure your text (letters) are black (or very! dark grey at least) and your background is white. That is what the engine was trained on and for and thus black text on white background is what you should strive for in your images which you intend to feed to tesseract.

(See also notes in my email response in another thread in here just a few minutes ago. It in the documentation, but when you dont realize what youre reading there, this is the main thing to check and ensure to have:

- text is black

- background is white

- greyscale image is fine and possibly better for the added edge detail, but you invariably aim for something that looks as close as possible to "black print on white paper".

This the shown output (processed) image should be inverted to match the above conditions.

HTH