0

I'm trying to fetch score info from a Dota2 screenshot (disregard the "wrong" boxes on the borders)

original image

but I can't seem to get a good enough accuracy

after tesseract I'm applying these filters to the image

bw_image = cv2.bitwise_not(img)
bw_image = cv2.cvtColor(bw_image, cv2.COLOR_BGR2GRAY)

and this config to tesseract config = ("-l eng --oem 1 --psm 7")

In the screenshot provided the ROI was resized to (50,50) but resizing doesn't impact accuracy consistently

How can I improve my results? I searched how to train/improve a tesseract model but it seems to much work for what I want.

Carlos Silva
  • 101
  • 2
  • 9
  • as I remeber `tesseract` was created for scaners which scan white papers with black text, and rather not for images. It doesn't know if you can train it - it may not use machine learning or neural network. – furas Sep 12 '19 at 17:05
  • I found information about [training tesseract](https://github.com/tesseract-ocr/tesseract/wiki/TrainingTesseract) but it seems it needs long work. Maybe first try to crop image to keep only region with value and make more image modification - like change constrast - to create better image for OCR. See [Improving the quality of the output](https://github.com/tesseract-ocr/tesseract/wiki/ImproveQuality) – furas Sep 12 '19 at 17:15

0 Answers0