Image Preprocessing for OCR - Tessaract

Question

Obviously this image is pretty tough as it is low clarity and is not a real word. However, with this code, I'm detecting nothing close:

import pytesseract
from PIL import Image, ImageEnhance, ImageFilter
image_name = 'NedNoodleArms.jpg'
im = Image.open(image_name) 
im = im.filter(ImageFilter.MedianFilter())
enhancer = ImageEnhance.Contrast(im)
im = enhancer.enhance(2)
im = im.convert('1')
im.save(image_name)
text = pytesseract.image_to_string(Image.open(image_name))
print(text)

outputs

, Mdfiaodfiamms

Any ideas here? The image my contrasting function produces is:

Which looks decent? I don't have a ton of OCR experience. What preprocessing would you recommend here? I've tried resizing the image larger, which helps a little bit but not enough, along with a bunch of different filters from PIL. Nothing getting particularly close though

Lithopone

Do not convert to 1 bit B/W, use grayscale ('L" IINM). – Gourd

Thanks Paulo! That helped a lot. Its outputting 'NedNnodleArrns', which is super reasonable – Lithopone

Glad to help, check my answer. – Gourd

Answer 1

You are right, tesseract works better with higher resolutions so sometimes resizing the image helps - but don't convert to 1 bit.

I got good results converting to grayscale, making it 3 times as large and making the letters a bit brighter:

>>> im = Image.open('j78TY.png')\
          .convert('L').resize([3 * _ for _ in im.size], Image.BICUBIC)\
          .point(lambda p: p > 75 and p + 100)
>>> pytesseract.image_to_string(im)
'NedNoodleArms'

Check this jupyter notebook:

Gourd

Can you explain what this lambda function is doing? Is it like thresholding? – Swoosh

@Swoosh yeah, It is a lame hack for making the letters brighter, if a pixel value is over 75 (of 256) then add 100 to its value. – Gourd

You mean (of 255), right? What if pixel values is already over 200? – Swoosh

Awesome answer for an awesomely effective simple trick! Thanks! – Bertram