
                         [INLINE] [INLINE] [INLINE]
                                      
   Dsormais, nul n'osa pntrer dans la fort sauvage: l'effroi en garde
   l'entre et les amants y sont matres. C'est alors que Tristan faonna
   l'arc Qui-ne-faut, lequel atteignait toujours le but, homme ou bte, 
   l'endroit vis [Le roman de Tristan et Iseult, traduction Joseph
   Bdier]
   
What is it?

   Qui-ne-faut is an Optical Character Recognition (OCR) filter
   distributed under GNU General Public License. Although its name may
   accredit the thought it doeth not fail, it will only succeed in 70%
   correct recognition. For the while. I hope.
   
   Do not underestimate the power of the GNU GPL. It is neither public
   domain, nor freeware,... read this
   
Features

  Input
  
     * a raw grayscale PNM file: toto.pnm (this is the only format
       supported for the while, no ASCII PNM please).
     * must be a properly cropped page of text with first lines at top
       and last ones at bottom
     * Qui-ne-faut accepts, detects and corrects a skew = a slight
       rotation error on the page, up to 30 degrees (if you scan 30
       degrees wrong, you're clearly drunk).
     * For the while, no accents
       
  Output
  
     * an ASCII text
     * is it necessary to say that this output should instantly be piped
       into ispell to try automatic orthographic correction?
       
Features: details

     * Today's scanner allow at least 300dpi *true resolution* scans,
       which gives a very good basis to analyze characters very closely
       (order of magnitude: a character is grosso modo a 50x30 bitmap).
       When I suspect traditional software to use only lineart (that is,
       full black or full white pixel scans), quinefaut works in
       *grayscale* (that is, 255 levels of gray), to rebuild the outlines
       of characters and feed the classification tree with unusual
       classifiers.
     * text must be cleanly printed, with rectilign and parallel lines.
       Lines do not have to be regularly spaced.
     * The less the initial skew is, the better it can be corrected.
       Correction manages in most cases to guess the rotation to apply on
       the image at a precision of 0.001 rad, that is, 1,2, or 3 pixel
       residual deviation for a page that is 1000 pixels wide.
       
Qui-ne-faut FAQ

Installation and usage instructions

Download

     * Source (40k)
     * Documentation (2.5M)
     * Samples (8.2M)
       
                                  [INLINE]
