Extracting code from photograph of T-shirt via OCR

I recently saw someone with a T-shirt with some Perl code on the back. I took 开发者_StackOverflow中文版a photograph of it and cropped out the code:

Next I tried to extract the code from the image via OCR, so I installed Tesseract OCR and the Python bindings for it, pytesser.

Pytesser only works on TIFF images, so I converted the image in Gimp and entered the following code (Ubuntu 9.10):

>>> from pytesser import *
>>> image = Image.open('code.tif')
>>> print image_to_string(image)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "pytesser.py", line 30, in image_to_string
    util.image_to_scratch(im, scratch_image_name)
  File "util.py", line 7, in image_to_scratch
    im.save(scratch_image_name, dpi=(200,200))
  File "/usr/lib/python2.6/dist-packages/PIL/Image.py", line 1406, in save
    save_handler(self, fp, filename)
  File "/usr/lib/python2.6/dist-packages/PIL/BmpImagePlugin.py", line 197, in _save
    raise IOError("cannot write mode %s as BMP" % im.mode)
IOError: cannot write mode RGBA as BMP
>>> r,g,b,a = image.split()
>>> img = Image.merge("RGB", (r,g,b))
>>> print image_to_string(img)
Tesseract Open Source OCR Engine

     éi     _   l_` _ t  
  ’   ‘" fY`  
  {  W       IKQW
  ·  __·_  ‘ ·-»·      
       :W   Z  
  ··  I  A n   1   
           ;f        
     `    `      
`T     .' V   _ ‘  
I  {Z.; » ;,. , ;  y i-   4 : %:,,    
      `· »    V; ` ?    
‘,—·.    
H***li¥v·•·}I§¢   ` _  »¢is5#__·¤G$++}§;“»‘7·
  71   ’    Q  {  NH IQ
  ytéggygi {     ;g¤qg;gm·;,g(g,,3) {3;;+-
   § {Jf**$d$ }‘$p•¢L#d¤ Sc}
  »   i `  i A1:

That's clearly gibberish that comes out of the OCR engine. So, my question is:

• What do I have to do to get better OCR results out of Tesseract?
• Or, does anybody else have better luck extracting the code from the above image in another way?

You can probably type faster than you can clean up images and install OCR engines:

#!/usr/bin/perl
(my$d=q[AA                GTCAGTTCCT
  CGCTATGTA                 ACACACACCA
    TTTGTGAGT                ATGTAACATA
      CTCGCTGGC              TATGTCAGAC
        AGATTGATC          GATCGATAGA
          ATGATAGATC     GAACGAGTGA
            TAGATAGAGT GATAGATAGA
              GAGAGA GATAGAACGA
                TC GATAGAGAGA
                 TAGATAGACA G
               ATCGAGAGAC AGATA
             GAACGACAGA TAGATAGAT
           TGAGTGATAG    ACTGAGAGAT
         AGATAGATTG        ATAGATAGAT
       AGATAGATAG           ACTGATAGAT
     AGAGTGATAG             ATAGAATGAG
   AGATAGACAG               ACAGACAGAT
  AGATAGACAG               AGAGACAGAT
  TGATAGATAG             ATAGATAGAT
  TGATAGATAG           AATGATAGAT
   AGATTGAGTG        ACAGATCGAT
     AGAACCTTTCT   CAGTAACAGT
       CTTTCTCGC TGGCTTGCTT
         TCTAA CAACCTTACT
           G ACTGCCTTTC
           TGAGATAGAT CGA
         TAGATAGATA GACAGAC
       AGATAGATAG  ATAGAATGAC
     AGACAGAGAG      ACAGAATGAT
   CGAGAGACAG          ATAGATAGAT
  AGAATGATAG             ACAGATAGAC
  AGATAGATAG               ACAGACAGAT
  AGACAGACTG                 ATAGATAGAT
   AGATAGATAG                 AATGACAGAT
     CGATTGAATG               ACAGATAGAT
       CGACAGATAG             ATAGACAGAT
         AGAGTGATAG          ATTGATCGAC
           TGATTGATAG      ACTGATTGAT
             AGACAGATAG  AGTGACAGAT
               CGACAGA TAGATAGATA
                 GATA GATAGATAG
                    ATAGACAGA G
                  AGATAGATAG ACA
                GTCGCAAGTTC GCTCACA
])=~s/\s+//g;%a=map{chr $_=>$i++}65,84,67,
71;$p=join$;,keys%a;while($d=~/([$p]{4})/g
){next if$j++%96>=16;$c=0;for$d(0..3){$c+=
$a{substr($1,$d,1)}*(4**$d)}$perl.=chr $c}
             eval $perl;

Edit: typo.

pre-processing will definitely yield a more workable image.

For example, here is the result of Gimp "Levels", "Difference-of-Gaussians", and "Levels" filters on the image.

Just a few small typos in RedDwight code.

#!/usr/bin/perl
(my $d=q[AA                GTCAGTTCCT
  CGCTATGTA                 ACACACACCA
    TTTGTGAGT                ATGTAACATA
      CTCGCTGGC              TATGTCAGAC
        AGATTGATC          GATCGATAGA
          ATGATAGATC     GAACGAGTGA
            TAGATAGAGT GATAGATAGA
              GAGAGA GATAGAACGA
                TC GATAGAGAGA
                 TAGATAGACA G
               ATCGAGAGAC AGATA
             GAACGACAGA TAGATAGAT
           TGAGTGATAG    ACTGAGAGAT
         AGATAGATTG        ATAGATAGAT
       AGATAGATAG           ACTGATAGAT
     AGAGTGATAG             ATAGAATGAG
   AGATAGACAG               ACAGACAGAT
  AGATAGACAG               AGAGACAGAT
  TGATAGATAG             ATAGATAGAT
  TGATAGATAG           AATGATAGAT
   AGATTGAGTG        ACAGATCGAT
     AGAACCTTTCT   CAGTAACAGT
       CTTTCTCGC TGGCTTGCTT
         TCTAA CAACCTTACT
           G ACTGCCTTTC
           TGAGATAGAT CGA
         TAGATAGATA GACAGAC
       AGATAGATAG  ATAGAATGAC
     AGACAGAGAG      ACAGAATGAT
   CGAGAGACAG          ATAGATAGAT
  AGAATGATAG             ACAGATAGAC
  AGATAGATAG               ACAGACAGAT
  AGACAGACTG                 ATAGATAGAT
   AGATAGATAG                 AATGACAGAT
     CGATTGAATG               ACAGATAGAT
       CGACAGATAG             ATAGACAGAT
         AGAGTGATAG          ATTGATCGAC
           TGATTGATAG      ACTGATTGAT
             AGACAGATAG  AGTGACAGAT
               CGACAGA TAGATAGATA
                 GATA GATAGATAG
                    ATAGACAGA G
                  AGATAGATAG ACA
                GTCGCAAGTTC GCTCACA
])=~s/\s+//g;%a=map{chr $_=>$i++}65,84,67,
71;$p=join$;,keys%a;while($d=~/([$p]{4})/g
){next if$j++%96>=16;$c=0;for$d(0..3){$c+=
$a{substr($1,$d,1)}*(4**$d)}$perl.=chr $c}
             eval $perl;

that when executed produces:

Just another genome hacker.

If I were you I'd start by cleaning up the image as much as possible, using a picture-manipulation program (GIMP, for example) so that the input for the OCR would be more easily understandable.

If possible, aim for creating a black-and-white only image.

Hmm perhaps you need to process the image, i.e. put it though some filters like 'edge detection', emboss/engrave or a noise filter...

Good OCRs are strongly guided by redundancies in natural languages to yield a subset for "what might be the next character". Perl code gives no such aid to the OCR. Type it in by hand.

The key for a task like this is to take advantage of the evident constraints. Find a library that lets you specify your own character set. Require all the characters in the main DNA helices to be one of A T G C. Require that the whole thing parse as perl. Type in the hard parts by hand if necessary.