1

I have an image with mostly light text on a dark background. Some of the text is a darker color (purple).

I'm using opencv-python to manipulate the image for better OCR parsing.

There is a little more processing that happens before this, but I feel like the processing steps giving me trouble are as follow.

The image gets converted to grayscale

enter image description here

cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)

The image then gets inverted (this seems to keep the final text clearer)

enter image description here

cv2.bitwise_not(img)

The image then gets run through threhold

enter image description here

cv2.threshold(img, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)[1]

You can see I'm totally losing the darker text. Switching to an adaptive threshold does preserve the text better but creates a ton of noise (the background appears flat black but is not).

Any thoughts on how I can modify my current thresholding to preserve that darker text?

Christoph Rackwitz
  • 11,317
  • 4
  • 27
  • 36
Bryant Makes Programs
  • 1,493
  • 2
  • 17
  • 39

1 Answers1

1

It appears your use case isn't ideal for Otsu, I'd remove the option from threshold and hardcode or calculate through other method the threshold range.

Hardcoded to these values:

cv2.threshold(img, 215, 255, cv2.THRESH_BINARY)[1]

Yields the following result:

enter image description here

Calculating though other method would depend on the range of possible inputs you have. Depending on the range of possible inputs you might want to filter or preprocess the image in other ways before the thresholding.

Delta de Dirac
  • 339
  • 5