Tesseract not recognizing digits correctly from OpenCV preprocessing

Question

I'm trying to OCR these digits. However, tesseract is not recognizing them properly.

import cv2 from pytesseract import image_to_string

image = cv2.imread('PATHTOIMAGE', cv2.IMREAD_COLOR)
image = cv2.resize(image, None, fx=5, fy=5, interpolation=cv2.INTER_CUBIC)
gaussian = cv2.GaussianBlur(image, (5, 5), 2)
mask = cv2.inRange(gaussian, (250, 250, 250), (255, 255, 255))
ocr = image_to_string(mask, config='-c tessedit_char_whitelist=0123456789')
print(ocr)

The masking result is the following:

OCR result: 88311

I tried performing some morphological operations from here (dilating and opening), but no luck.

I also tried to detect contours and detect digit by digit, but also no luck.

How else could I improve?

You may try some of the --psm options, such as --psm 5 muthu.co/all-tesseract-ocr-options — Davis
– Davis, Commented Jan 10, 2022 at 21:14
Awesome. However, it's confusing 2 with 8. Any preprocessing suggestion for that? — André Luís
– André Luís, Commented Jan 11, 2022 at 15:16

Davis · Accepted Answer · 2022-01-11 17:51:18Z

1

I was able to achieve correct results without the scaling and Gaussian blur steps. I also inverted the mask and used only --psm 6

image = cv2.imread('PATHTOIMAGE', cv2.IMREAD_COLOR)
mask = cv2.inRange(image, (250, 250, 250), (255, 255, 255))
ocr = image_to_string(~mask, config='--psm 6')
print(ocr)

For what it's worth, I cannot recreate the 2 vs 8 confusion error. I am running tesseract 4.1.1 and pytesseract 0.3.8 on windows.

If you still need to explore other preprocessing steps, consider running a sequence of erosion and dilation operations with asymmetric kernels. For example

kernel = np.ones((6,2), np.uint8)
img = cv2.erode(mask, kernel, 1)

kernel = np.ones((2,40), np.uint8)
img = cv2.dilate(img, kernel, 1)

kernel = np.ones((1,40), np.uint8)
img = cv2.erode(img, kernel, 1)

kernel = np.ones((4,2), np.uint8)
img = cv2.dilate(img, kernel, 1)

Numbers are for your scaled image. This appears to remove the vertical points on the 2 while retaining other relevant information. You will need to tune and ensure it works for all number options.

answered Jan 11, 2022 at 17:51

Davis

6126 silver badges13 bronze badges

Sign up to request clarification or add additional context in comments.

1 Comment

bfris Over a year ago

Bravo. Well done. I played with this myself and couldn't get it to work. It seems the scaling in the OP is what messes up the recognition. Anything above 2x scaling seems to not work.

Collectives™ on Stack Overflow

Tesseract not recognizing digits correctly from OpenCV preprocessing

1 Answer 1

1 Comment

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

1 Comment

Your Answer

Sign up or log in

Post as a guest

Related