Train tesseract 3 to get table of letters
image-processing, image-recognition, ocr, pattern-recognition, tesseract
Solution
First, I suggest you preprocess your image, for example making the dark parts darker, blur it a little. Feel free to experiment until Tesseract stops seeing letters in the filled-in squares.
Second, you have two options:
One, you can enable hOCR output and try to parse the layout of the scanned letters yourself. hOCR is a subset of HTML and it contains coordinates of all recognized words. Try figuring out where the rows and columns are.
Alternatively, try making Tesseract recognise the layout properly, not rotated 90°.
Anyway, this is what I did:
1. I ran the image through ImageMagick:
` $ convert CDZjN.png -deskew 40% -contrast-stretch 7%x10% -filter lanczos -resize 250% ooo.png `
2. I created a config file `t.conf` for Tesseract, disabling vertical text detection and English dictionary:
` textord_tabfind_vertical_text 0 load_system_dawg 0 load_freq_dawg 0 load_punc_dawg 0 load_number_dawg 0 load_unambig_dawg 0 load_bigram_dawg 0 load_fixed_length_dawgs 0 `
3. I simply ran it:
` $ tesseract ooo.png ooo t.conf ; cat ooo.txt Tesseract Open Source OCR Engine v3.02 with Leptonica 01ABC-E 26ABCDE 02A CDE 27ABCDE o3 BCDE 28ABCDE o4 BCDE 29ABCDE o5 BCDE 30ABCDE 06ABCD. 31ABCDE 07A-CDE 32ABCDE 08ABC.E 33ABCDE o9 BCDE 34ABCDE 10A CDE 35ABCDE 11ABCD 36ABCDE 12ABC E 37ABCDE 13ABC E 38ABCDE 14ABCD 39ABCDE 15 BCDE 40ABCDE 1s BCDE 41ABCDE 17 BCDE 42ABCDE 18ABCD_ 43ABCDE 19AB DE 44ABCDE 20AB DE 45ABCDE 21ABCDE 46ABCDE 22ABCDE 47ABCDE 23ABCDE 48ABCDE 24ABCDE 49ABCDE 25ABCDE 50ABCDE `
Not perfect, but passable.
Problem
I've been trying to use plain tesseract 3 OCR using different options to get the data from a table of letters where my students marked one as answers for multiple choice questions, as seen below: One of the best outputs was: ``` EEEEEEEEEEEEEEEEEEEEEEEEE DDDDDDDDDDDDDDDDDDDDDDDDD CCCCCCCCCCCCCCCCCCCCCCCCC BBBBBBBEBBBBBBBBBBBBBBBBB AAAAAAAAAAAAAAAAAAAAAAAAA 6789012345678901234567890 2222333333333344444444445 EEEEE EEEE EE EEE EEEEEEE DDDDDD DDD DDDDDDDDDDDD CCCCCCCCCCCCCCCCCC CCCCC B BEBE BB BBBBBBBBBBBBBBB AA AAA AAAAA AAAAAAAA 1234567890123455789012345 OOOOOOOOO1111111111222222 ``` I know I can parse that .txt and have a better result, but it missed a lot of information and got the letters from some of the painted blocks. I wanted to know what can I do to get better result for this case. I would also like to have a table with the painted blocks appearing as a different character, for example, for the first and second lines of the image: ``` 01 A B C - E 26 A B C D E 02 A - C D E 27 A B C D E ``` If you guys have some similar experience, any information will be appreciated! Thanks in advance!