tesseract fails at simple number detection - ocr

Question

I want to perform OCR on images like this one:
It is a table with numerical data with colons as decimal separators.
It is not noisy, contrast is good, black text on white background.
As an additional preprocessing step, in order to get around issues with the frame borders, I cut out every cell, binarize it, pad it with a white border (to prevent edge issues) and pass only that single cell image to tesseract.
I also looked at the individual cell images to make sure the cutting process works as expected and does not produce artifacts. These are two examples of the input images for tesseract:
Unfortunately, tesseract is unable to parse these consistently. I have found no configuration where all 36 values are recognized correctly.
There exist a couple similar questions here on SO and the usual answer is a suggestion for a specific combination of the --oem and --psm parameters. So I wrote a python script with pytesseract that loops over all combinations of --oem from 0 to 3 and all values of --psm from 0 to 13 as well als lang=eng and lang=deu. I ignored the combinations that throw errors.
Example 1: With --psm 13 --oem 3 the above "1,7" image is misidentified as "4,7", but the "57" image is correctly recognized as "57".
Example 2: With --psm 6 --oem 3 the above "1,7" image is correctly recognized as "1,7", but the "57" image is misidentified as "o/".
Any suggestions what else might be helpful in improving the output quality of tesseract here?
My tesseract version:
tesseract v4.0.0.20190314
leptonica-1.78.0
libgif 5.1.4 : libjpeg 8d (libjpeg-turbo 1.5.3) : libpng 1.6.34 : libtiff 4.0.9 : zlib 1.2.11 : libwebp 0.6.1 : libopenjp2 2.2.0
Found AVX2
Found AVX
Found SSE

Answer 1

Solution
Divide the image into the 5-different row
Apply division-normalization to each row
Set psm to 6 (Assume a single uniform block of text.)
Read
From the original image, we can see there are 5 different rows.
Each iteration, we will take a row, apply normalization and read.
We need to understand how to set image indexes carefully.
import cv2
from pytesseract import image_to_string
img = cv2.imread("0VXIY.jpg")
gry = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
(h, w) = gry.shape[:2]
start_index = 0
end_index = int(h/5)
Question Why do we declare start and end indexes?
We want to read a single row in each iteration. Lets see in an example:
The current image height and width are 645 and 1597 pixels.
Divide the images based on indexes:
start-index
end-index
0
129
129
258 (129 + 129)
258
387 (258 + 129)
387
516 (387 + 129)
Lets see whether the images are readable?
start-index
end-index
image
0
129
129
258
258
387
387
516
Nope, they are not suitable for reading, maybe a little adjustment might help us. Like:
start-index
end-index
image
0
129 - 20
109
218
218
327
327
436
436
545
545
654
Now they are suitable for reading.
When we apply the division-normalization to each row:
start-index
end-index
image
0
109
109
218
218
327
327
436
436
545
545
654
Now when we read:
1,7 | 57 | 71 | 59 | .70 | 65
| 57 | 1,5 | 71 | 59 | 70 | 65
| 71 | 59 | 1,3 | 57 | 70 | 60
| 71 | 59 | 56 | 1,3 | 70 | 60
| 72 | 66 | 71 | 59 | 1,2 | 56
| 72 | 66 | 71 | 59 | 56 | 4,3
Code:
import cv2
from pytesseract import image_to_string
img = cv2.imread("0VXIY.jpg")
gry = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
(h, w) = gry.shape[:2]
# print(img.shape[:2])
start_index = 0
end_index = int(h/5) - 20
for i in range(0, 6):
# print("{}->{}".format(start_index, end_index))
gry_crp = gry[start_index:end_index, 0:w]
blr = cv2.GaussianBlur(gry_crp, (145, 145), 0)
div = cv2.divide(gry_crp, blr, scale=192)
txt = image_to_string(div, config="--psm 6")
print(txt)
start_index = end_index
end_index = start_index + int(h/5) - 20