Unable to identify image for number 5 using Eng trained data

62 views
Skip to first unread message

vinayb...@gmail.com

unread,
Oct 27, 2018, 2:52:45 AM10/27/18
to tesseract-ocr

I am using Pytesseract to recognise an image for number 5 and I'm stunned that even after applying various filters like GlaussianBlur and Threshold and applying dilation and erosion to remove the noise it still not able to identify the image.


I am using Eng Trained data by default. Not sure where I am going wrong. Do I need to include any other training file here?


Filters Tried:


          1: cv2.threshold(cv2.GaussianBlur(img, (9, 9), 0), 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)[1],

            2: cv2.threshold(cv2.GaussianBlur(img, (7, 7), 0), 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)[1],

            3: cv2.threshold(cv2.GaussianBlur(img, (5, 5), 0), 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)[1],

            4: cv2.threshold(cv2.medianBlur(img, 5), 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)[1],

            5: cv2.threshold(cv2.medianBlur(img, 3), 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)[1],

            6: cv2.adaptiveThreshold(cv2.GaussianBlur(img, (5, 5), 0), 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY, 31, 2),

            7: cv2.adaptiveThreshold(cv2.medianBlur(img, 3), 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY, 31, 2),



Training Data:

eng.traineddata


Original Image: See Attached





five_filter_5.jpg
five_filter_7.jpg

Vinod Gattani

unread,
Oct 27, 2018, 4:48:37 AM10/27/18
to tesser...@googlegroups.com

I used this command:

 tesseract five_filter_5.jpg ocr.txt --oem 1 --psm 6 -l eng

I used "eng.traineddata" from tessdata_best repo.

It gave "5" in ocr.txt.




--
You received this message because you are subscribed to the Google Groups "tesseract-ocr" group.
To unsubscribe from this group and stop receiving emails from it, send an email to tesseract-oc...@googlegroups.com.
To post to this group, send email to tesser...@googlegroups.com.
Visit this group at https://groups.google.com/group/tesseract-ocr.
To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/ec25b1e1-c9f3-4743-b2fd-6efdd2a978f6%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

vinayb...@gmail.com

unread,
Oct 27, 2018, 6:16:04 AM10/27/18
to tesseract-ocr
Can you try this new attached image for Alphabet "M" ?
m.PNG

Vinod Gattani

unread,
Oct 27, 2018, 7:52:34 AM10/27/18
to tesser...@googlegroups.com
It gave "|" as text.

When resized to 50*50, text is "N\". You should check whether font used in the image, is a part of fonts on which English language was trained.

Thanks

--
You received this message because you are subscribed to the Google Groups "tesseract-ocr" group.
To unsubscribe from this group and stop receiving emails from it, send an email to tesseract-oc...@googlegroups.com.
To post to this group, send email to tesser...@googlegroups.com.
Visit this group at https://groups.google.com/group/tesseract-ocr.

Vinay Babu

unread,
Oct 27, 2018, 8:12:14 AM10/27/18
to tesser...@googlegroups.com
Well it doesn't seems to be a problem with fonts Training. I tried capturing the same image without skewness and it perfectly worked out. Not sure why tesseract doesn't works with bit skewed texts in images..

vinayb...@gmail.com

unread,
Oct 27, 2018, 8:15:44 AM10/27/18
to tesseract-ocr
Here is another image where text is skewed and tesseract fails to identify it.


On Saturday, October 27, 2018 at 12:22:45 PM UTC+5:30, vinayb...@gmail.com wrote:
mericarclean1.PNG
Reply all
Reply to author
Forward
0 new messages