Page 1 of 1
Trying to get the best result with imagemagic and tesseract OCR, image recognition
Posted: 2017-07-04T19:52:59-07:00
by metal0077
Hello I'm trying to use OCR tesseract to recognize some letters in a image.
I did a convert using imagemagick and image seems to be good but its not enough to recognize
The original images:
The command used with imagemagick to convert(with a lot of thanks to @fmw42)
Code: Select all
convert input.jpg -fuzz 50% -fill black -opaque black -bordercolor white -border 2 -fill black -draw "color 0,0 floodfill" -alpha off -negate -units pixelsperinch -density 72 output.jpg
The result images:
The OCR tesseract command:
Output/result:
Text: AUGU -> AUOU
Tesseract Open Source OCR Engine v4.00.00alpha with Leptonica Page 1
Text: VEGU -> VOR-OU
Tesseract Open Source OCR Engine v4.00.00alpha with Leptonica Page 1
Text: EGUV -> E6UV
Tesseract Open Source OCR Engine v4.00.00alpha with Leptonica Page 1
Text: USEA -> USSOEA
May I can use some filter to get better results? what do you think?
Re: Trying to get the best result with imagemagic and tesseract OCR, image recognition
Posted: 2017-07-04T21:03:11-07:00
by fmw42
Your problem is really with your input quality. You should never use JPG since it is a lossy compression. You need to have scans at higher resolutions/density to get better quality results. Also I suspect the main issue is that your characters are rotated and I think OCR does not like rotated letters or numbers. The quality should have been adequate.
Re: Trying to get the best result with imagemagic and tesseract OCR, image recognition
Posted: 2017-07-04T23:00:09-07:00
by metal0077
fmw42 wrote: ↑2017-07-04T21:03:11-07:00
Your problem is really with your input quality. You should never use JPG since it is a lossy compression. You need to have scans at higher resolutions/density to get better quality results. Also
I suspect the main issue is that your characters are rotated and I think OCR does not like rotated letters or numbers. The quality should have been adequate.
I converted jpg to tiff before processing...
Code: Select all
convert input.jpg -units pixelsperinch -density 300 -depth 8 output.tiff
The result's change a little be but without correct..
Text: AUGU -> AUOU
Tesseract Open Source OCR Engine v4.00.00alpha with Leptonica
Page 1
Text: VEGU -> VOR°OU
Tesseract Open Source OCR Engine v4.00.00alpha with Leptonica
Page 1
Text: EGUV -> E6UV
Tesseract Open Source OCR Engine v4.00.00alpha with Leptonica
Page 1
Text: USEA -> USSOEA
Re: Trying to get the best result with imagemagic and tesseract OCR, image recognition
Posted: 2017-07-04T23:39:14-07:00
by fmw42
Once you have compressed the image using JPG going to tif will not help. You have already lost information. If you scanned these image, you would have been better of saving as tif or pdf.
But the real issue I believe is that the characters are rotated.
Re: Trying to get the best result with imagemagic and tesseract OCR, image recognition
Posted: 2017-08-10T18:51:00-07:00
by mathieularose
I tried with your images and it looks good:
Code: Select all
$ tesseract -psm 8 augu.jpg -
Ausu
With whitelisting:
Code: Select all
$ tesseract -psm 8 -c tessedit_char_whitelist=ABCDEFGHIJKLMOPQRSTUVWXYZ augu.jpg -
AUGU
Code: Select all
$ tesseract -psm 8 vegu.jpg -
VEGU
Code: Select all
$ tesseract -psm 8 eguv.jpg -
EGUV
Code: Select all
$ tesseract -psm 8 usea.jpg -
USEA
Here is my tesseract version:
Code: Select all
$ tesseract --version
tesseract 3.04.01
leptonica-1.74.1
libgif 4.1.6(?) : libjpeg 6b (libjpeg-turbo 1.5.1) : libpng 1.6.27 : libtiff 4.0.7 : zlib 1.2.8 : libwebp 0.5.2
I posted your example at
https://mathieularose.com/decoding-captchas/