3dpcp/.svn/pristine/a8/a8681ae31c7559fca0d139c34479e6bf664ac771.svn-base

578 lines
21 KiB
Text
Raw Normal View History

2012-09-16 12:33:11 +00:00
<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 3.2 Final//EN">
<HTML>
<HEAD>
<TITLE>GOCR-documentation</TITLE>
<META NAME="description" CONTENT="GOCR-documentation">
<META NAME="keywords" CONTENT="ocr">
<META NAME="resource-type" CONTENT="document">
<META NAME="distribution" CONTENT="global">
<META HTTP-EQUIV="Content-Type" CONTENT="text/html; charset=iso-8859-1">
<META NAME="Generator" CONTENT="Joerg">
</HEAD>
<BODY >
<H1 ALIGN="CENTER">GOCR-documentation</H1>
<P ALIGN="CENTER"><STRONG>J&ouml;rg Schulenburg</STRONG></P>
<P ALIGN="CENTER"><STRONG>Magdeburg, June 3, 2002</STRONG></P>
<H3>Abstract:</H3>
<DIV>
In this documentation I describe some ideas for my OCR-program.
It contains algorithms and examples and gives you
an impression of what the program can (or could) do.
</DIV>
<P>
<P>
<BR><HR>
<!--Table of Child-Links-->
<A NAME="CHILD_LINKS"></A>
<UL>
<LI><A HREF="#SEC1">Introduction</A>
<LI><A HREF="#SEC2">Segmentation of textual regions / Layout analysis</A>
<LI><A HREF="#SEC3">Line detection</A>
<LI><A HREF="#SEC4">Cluster detection</A>
<LI><A HREF="#SEC5">Engines</A>
<LI><A HREF="#SEC6">Remove pixels</A>
<LI><A HREF="#SEC7">Add pixels</A>
<LI><A HREF="#SEC8">Similarity analyzer</A>
<LI><A HREF="#SEC9">Overlapping characters</A>
<LI><A HREF="#SEC10">Black/White, Gray and Colors</A>
<LI><A HREF="#SEC11">Pictures on scanned pages</A>
<LI><A HREF="#SEC12">Tools</A>
<LI><A HREF="#SEC13">glossary</A>
<LI><A HREF="#SEC14">More information?</A>
<LI><A HREF="#SEC15">About this document</A>
</UL>
<!--End of Table of Child-Links-->
<H1><A NAME="SEC1"> Introduction</A>
</H1>
First I have to say that I am not a expert in pattern recognition
or similar things. My knowledge is based mostly on experiments with my
program.
Therefore do not worry about stupid algorithms I put in this document.
In this documentation I describe some ideas for my OCR-program.
The examples give you an impression of how the program handles
your images.
If you have comments regarding contents or spelling please
write to the author.
<H1><A NAME="SEC2">Segmentation of textual regions / Layout analysis</A></H1>
This is implemented as a recursive division in two parts.
<UL>
<LI>look for the thickest horizontal or vertical gap through the box</LI>
<LI>if the gap is less than five times longer than thick do not divide </LI>
<LI>do the same with the two new parts</LI>
</UL>
I know that this algorithm is not as good as you wish,
but I do not know a better one.
<P>
It would be very helpful to know about a function which is able to
decide whether the box represents a single text line or a more complex object.
<H1><A NAME="SEC3">Line detection</A></H1>
<P>
Line detection is very importand for good recognition.
For example it is difficult to distinguish between lowercase letter <B>p</B>
and uppercase letter <B>P</B> without having a baseline (same total height).
The lowercase version of <B>p</B> has a depht (the lower end is below the
baseline) and therefore its easy to distinguish from the uppercase version
if the baseline is known. The line detection is responsible for finding the
baseline of every text line.
<P>
Lines of characters are detected by looking for interline spaces.
These are characterized by a large number of non-black pixels in a
row. Image rotation (skewing) presents a problem, therefore the program
first looks only at the left half of the image. When a line is
found, the left half of the right side is scanned, because lines
are often short. The variation in height gives an indication of
the rotation angle. Using this angle, a second run detects lines
more accurately. Line detection may fail if there is dust on the
image.
<P>
In version v0.2.3 this behaviour is slightly changed.
To detect the rotation angle, the line through the most
characters is detected.
<H1><A NAME="SEC4">Cluster detection</A></H1>
A cluster is a group of pixels which are connected with each other.
The simplest way to detect a cluster is to look for a pixel.
If you find one, look to the neighbouring pixels. This can be done recursively.
<P>
This method needs a lot of stack space if a cluster is very large,
and can cause problems with the memory.
<P>
Do you remember the algorithm for leaving a maze?
Go along the right (or left) wall. This seems to be a good approach
for detecting clusters without recursion.
The following picture shows a trace of the maze algorithm.
<P>
<TABLE WIDTH="680">
<TR><TD>
<PRE>
first 35 steps next 36 steps
..@@@@@..@@@@&lt;.. ..v&lt;&lt;&lt;&lt;..v&lt;&lt;&lt;@.. * = starting point
..@@@@@@@@@.@^&lt;. ..&gt;&gt;v@^&lt;&lt;&lt;@.@@@. &gt;^&lt;v = go right,up,left,down
....@@@@@...@@^. ....v@@@@...@@@. @ = black pixel
....@@@@....@@^. ....v@@@....@@@.
....@@@.....@@^. ....v@@.....@@@.
....@@@.....@@^. ....v@@.....@@@.
...@@@@.....@@^. ...v&lt;@@.....@@@.
...@@@......@@^. ...v@@......@@@.
...@@@......@@^. ...v@@......@@@.
...@@@.....@@@^. ...v@@.....@@@@.
...@@@.....@@&gt;^. ...v@@.....@@@@.
...@@@.....@@^.. ...v@@.....@@@..
..@@@@.....@@^.. ..v&lt;@@.....@@@..
..@@@@....@@@^.. ..v@@@....@@@@..
*&gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt;&gt;^&lt;&lt; @@@@@@@@@@@@@@@@
</PRE></TD></TR>
</TABLE>
<P>
The minimum and maximum coordinates can be used to create a box around the
cluster. But does this algorithm work with diagonally connected pixels?
<H1><A NAME="SEC5">Engines</A></H1>
GOCR is able to work with different recognition engines.
Since version 0.37 engines have to return a probability value together
with the recognized character or a table of values to a table of characters.
If the probability value is 100, the engine is 100% sure to have found the
right character otherwise the value is less. This gives GOCR the possibility
to compare results of different engines or in case of a not recognized character
to inform the user or another
application (spell checker) which characters probably could be there.
<H2>Base-Engine</H2>
The base engine (src/ocrX.c) is the original engine used in the first implementation
of GOCR by J&ouml;rg. The idea was to get a fast and acceptable result
without learning theoretical background. Later it should be replaced or completed
by a better engine.
The base engine is a rule based engine.
The engine was written without theoretical
background and is improved by try and error method but is is still far from
perfect. The algorithm is very tolerant to size and form af characters
(omnifont).
How does the engine identify a character? For the explanation look at the
following pattern.
<P>
<TABLE WIDTH="680">
<TR><TD>
<PRE>
vvvv vv- white regions
......@@...... &lt;- crossing one line
......@@......
.....@@@@.....
.....@@@@.....
.....@@@@.....
....@..@@@.... &lt;- white hole / crossing two lines
....@..@@@.... &lt;- crossing two lines
....@..@@@....
...@....@@@...
...@....@@@...
...@....@@@...
..@@@@@@@@@@.. &lt;- horizontal line near center
..@......@@@..
..@......@@@..
.@........@@@. v- increasing width of pattern
.@........@@@. v
.@........@@@. v
@@@......@@@@@
^^^-- gap
</PRE></TD></TR>
</TABLE>
<P>
In the future the program
should detect edges, vertices, gaps, angles and so on.
This is called feature extraction (as far as I know).
With such data the engine could make a cluster analysis.
But this is a difficult task, if the scanned image is noisy.
<H2>Database-Engine</H2>
The database engine (src/database.c) was the second engine added to GOCR.
It was primary written to give users a simple tool to recognize
special language-specific characters. The program generates a list
(text file db.lst of image filenames and character codes)
and image samples (pnm-files) in a database path (./db/).
The database can be created by hand or extern programs or by GOCR itself
using option (-m 130). In the last case GOCR prompts the user
for not recognized characters. If he enters the character the pattern
is saved in the database path as pnm-file and its file name is added
to the database list (db.lst) together with the text string the pattern
should be replaced by.
For recognition GOCR first loads the database into memory (option -m 2).
The main algorithm compares not recognized characters with stored images
and calculates a distance value. If the distance value is small enough,
the character is treated as recognized.
<H1><A NAME="SEC6">Remove pixels</A></H1>
The following picture shows an <I>n</I> which has additional pixels at the
bottom. Therefore it can not be detected as <I>n</I>. What can be done?
<UL>
<LI>classify horizontal (<TT>=</TT>) and vertical (<TT>I</TT>) pixels by
comparing the distance between the next vertical and next horizontal white
pixels (.)
</LI>
<LI>measure mean thickness of vertical and horizontal clusters
</LI>
<LI>erase unusually thin horizontal pixels at the bottom line
</LI>
</UL>
<P>
<TABLE WIDTH="680">
<TR><TD>
<PRE>
..@@@@@..@@@@@.. ..==III..===II.. dx=16 dy=15
..@@@@@@@@@.@@@. ..==III====.III. thickness 2 to 3
....@@@@@...@@@. ....III==...III.
....@@@@....@@@. ....III=....III.
....@@@.....@@@. ....III.....III.
....@@@.....@@@. ....III.....III.
...@@@@.....@@@. ...IIII.....III.
...@@@......@@@. ...III......III.
...@@@......@@@. ...III......III.
...@@@.....@@@@. ...III.....IIII.
...@@@.....@@@@. ...III.....IIII.
...@@@.....@@@.. ...III.....III..
..@@@@.....@@@.. ..IIII.....III..
..@@@@....@@@@.. ..IIII....IIII..
@@@@@@@@@@@@@@@@ ================
^^^
this causes the problem
</PRE></TD></TR>
</TABLE>
<P>
A better way is to find serifs (horizontal lines glued on the lower end
of vertical lines) which touch together (v0.2.5).
<P>
The next picture shows blind pixels which are caused by dust on the paper.
The upper right dots are not connected with the rest of the character.
This can be detected via fill-algorithms. Currently the program
assumes that dots near the upper end of a character are ``i''-dots
or diaereses (umlaut dots).
<P>
<TABLE WIDTH="680">
<TR><TD>
<PRE>
..........................O... ..........................O...
..........................O... ..........................O...
.............................. ..............................
.............................. ..............................
..........@@@.......@@@@...... ..........@@@.......@@@@......
..@@@@..@@@@@@@...@@@@@@@..... ..@@@@..@@@@@@@...@@@@@@@.....
@@@@@@@@@@@@@@@@.@@@@@@@@@.... @@@@@@@@@@@@@@@@.@@@@@@@@@....
..@@@@@@....@@@@@@.....@@@@... ..@@@@@@....@@@@@@.....@@@@...
..@@@@.......@@@@......@@@@... ..@@@@.......@@@@......@@@@...
..@@@@.......@@@.......@@@@... ..@@@@.......@@@.......@@@@...
..@@@@.......@@@.......@@@@... ..@@@@.......@@@.......@@@@...
..@@@@.......@@@.......@@@@... ..@@@@.......@@@.......@@@@...
..@@@@.......@@@@......@@@@... ..@@@@.......@@@@......@@@@...
..@@@@.......@@@.......@@@@... ..@@@@.......@@@.......@@@@...
..@@@@.......@@@@......@@@@... ..@@@@.......@@@@......@@@@...
..@@@@.......@@@.......@@@@... ..@@@@.......@@@.......@@@@...
..@@@@.......@@@.......@@@@... ..@@@@.......@@@.......@@@@...
..@@@@.......@@@.......@@@@... ..@@@@.......@@@.......@@@@...
..@@@@.......@@@.......@@@@... ..@@@@.......@@@.......@@@@...
..@@@@.......@@@.......@@@@... ..@@@@.......@@@.......@@@@...
..@@@@.......@@@@......@@@@... ..@@@@.......@@@@......@@@@...
..@@@@......@@@@@......@@@@@.. ..@@@@......@@@@@......@@@@@..
@@@@@@@@..@@@@@@@@@..@@@@@@@@@ @@@@@@@@..@@@@@@@@@..@@@@@@@@@
</PRE></TD></TR>
</TABLE>
<H1><A NAME="SEC7">
Add pixels</A>
</H1>
The following picture shows an <I>m</I>. The legs are only barely connected.
How do we handle this?
<UL>
<LI>if the engine has failed, a filter is switched on and the engine
starts over
</LI>
<LI>the 2x2 filter sets pixels to (<I>O</I>) near barely connected pixels
</LI>
</UL>
<P>
<TABLE WIDTH="680">
<TR><TD>
<PRE>
vv vv
@@@.@@@..@@@... @@@.@@@..@@@...
.@@.@@@@.@@@@..&lt; .@@O@@@@O@@@@.. filter: .@ =&gt; O@ @. =&gt; @O
.@@@..@@@..@@..&lt; .@@@..@@@..@@.. @. =&gt; @. .@ =&gt; .@
.@@@..@@@..@@@. .@@@..@@@..@@@.
.@@@..@@@..@@@. .@@@..@@@..@@@.
.@@@..@@@..@@@. .@@@..@@@..@@@.
.@@@..@@@..@@@. .@@@..@@@..@@@.
.@@@..@@@..@@@. .@@@..@@@..@@@.
.@@@..@@@..@@@. .@@@..@@@..@@@.
.@@@..@@@..@@@. .@@@..@@@..@@@.
.@@@..@@@..@@@. .@@@..@@@..@@@.
.@@@..@@@..@@@. .@@@..@@@..@@@.
.@@@..@@@..@@@. .@@@..@@@..@@@.
@@@@@.@@@@.@@@@ @@@@@.@@@@.@@@@
</PRE></TD></TR>
</TABLE>
<H1><A NAME="SEC8">
Similarity analyzer</A>
</H1>
Some characters are a little bit noisy. These characters can be identified by
comparison with other, already recognized characters. This can be done
via a good distance function. May be
the distance function in the actual version of GOCR is not very good.
Feel free to send me your ideas, but be sure it does not waste my time.
<H1><A NAME="SEC9">
Overlapping characters</A>
</H1>
The following picture shows an overlapping <I>ru</I>.
How do we handle this?
<UL>
<LI>look for 3 weak connections (sum over y is small, start in the middle)
</LI>
<LI>test if the right and left part can be detected by the engine
</LI>
<LI>correction of surrounding box
</LI>
</UL>
<P>
<TABLE WIDTH="680">
<TR><TD>
<PRE>
....@@...@@@@@@@@@@....@@@@@@@.. ....@@...@@@@@@@@@@....@@@@@@@..
..@@@@..@@@@@..@@@@......@@@@@.. ..@@@@..@@@@@..@@@@......@@@@@..
@@@@@@@@@@@@@.,.@@@.......@@@@.. @@@@@@@@@@@@@...@@@.......@@@@..
..@@@@@@..@@@...@@@.......@@@@.. ..@@@@@@..@@@...@@@.......@@@@..
...@@@@.......,.@@@@......@@@@.. ...@@@@.........@@@@......@@@@..
...@@@@.........@@@@......@@@@.. ...@@@@.........@@@@......@@@@..
...@@@@.......,.@@@.......@@@@.. ...@@@@.........@@@.......@@@@..
...@@@@.........@@@.......@@@@.. ...@@@@.........@@@.......@@@@..
...@@@........,.@@@@......@@@@.. ...@@@..........@@@@......@@@@..
...@@@..........@@@@......@@@@.. ...@@@..........@@@@......@@@@..
...@@@........,.@@@@......@@@@.. ...@@@..........@@@@......@@@@..
...@@@..........@@@.......@@@@.. ...@@@..........@@@.......@@@@..
...@@@........,.@@@@......@@@@.. ...@@@..........@@@@......@@@@..
...@@@..........@@@@......@@@@.. ...@@@..........@@@@......@@@@..
...@@@........,.@@@@......@@@@.. ...@@@..........@@@@......@@@@..
...@@@..........@@@@@...@@@@@@@. ...@@@..........@@@@@...@@@@@@@.
..@@@@@.......,..@@@@@@@@@.@@@@@ ..@@@@@..........@@@@@@@@@.@@@@@
@@@@@@@@@.........@@@@@@@..@@@.. @@@@@@@@@.........@@@@@@@..@@@..
..............,....@@@.......... ...................@@@..........
^^^
213 weak vertical lines
</PRE></TD></TR>
</TABLE>
<P>
Of course the situation is more difficult with slanted characters.
<P>
The following example shows, how to deal with larger clusters.
To get a fast program a first test should select the possible positions of
division. That can be done by following upper and lower bows to a crease or a break. Than try to break on all detected creases, start at most
important one (not implemented yet v0.2.4).
<P>
<TABLE WIDTH="766">
<TR><TD>
<PRE>
&gt;&gt;&gt;&gt;vvv&lt;&lt;&lt;&lt;&lt; &gt;&gt;vv&lt;&lt;&lt;&lt; &gt;&gt;&gt;vvv&lt;&lt;&lt;&lt;
......@@@@@@@..................@@.........@@@@@@@..........@@@@@@@.....
....@@@@@@@@@@@...............@@@.......@@@@@@@@@@@......@@@@@@@@@@@...
...@@@@@@@@@@@@@.............@@@@......@@@@@@@@@@@@@....@@@@@@@@@@@@@..
..@@@@.......@@@@...........@@@@@.....@@@@.......@@@@..@@@@.......@@@@.
..@@@........@@@@..........@@@@@@@....@@@........@@@@@@@@@........@@@@.
.@@@@..........@@.........@@@@@@@@...@@@@..........@@@@@@@.........@@@@
.@@@.....................@@@@.@@@@...@@@..............@@...........@@@@
.@@@....................@@@@@.@@@@...@@@...........................@@@@
@@@...@@@@@@@...........@@@@..@@@...@@@...@@@@@@...................@@@.
@@@@.@@@@@@@@@@........@@@@...@@@@..@@@@.@@@@@@@@@@...............@@@@.
@@@@@@@@@@@@@@@.......@@@@....@@@@..@@@@@@@@@@@@@@@...............@@@..
@@@@@@@.....@@@@@.....@@@.....@@@@..@@@@@@......@@@@@............@@@@..
@@@@.........@@@@...@@@@......@@@@..@@@@@........@@@@...........@@@....
@@@@..........@@@@.@@@@.......@@@@..@@@@..........@@@..........@@@@....
@@@@..........@@@@@@@@@.......@@@@.@@@@@..........@@@.........@@@@.....
@@@@..........@@@@@@@@@@@@@@@@@@@@@@@@@@..........@@@@.......@@@@......
@@@@..........@@@@@@@@@@@@@@@@@@@@@@@@@@..........@@@@......@@@........
.@@@..........@@@@@@@@@@@@@@@@@@@@@@.@@@..........@@@@....@@@@@........
.@@@@........@@@@.............@@@@...@@@@........@@@@....@@@@..........
..@@@@.......@@@@.............@@@@....@@@@.......@@@@...@@@@...........
..@@@@@....@@@@@..............@@@@.....@@@@....@@@@@...@@@@@@..........
....@@@@@@@@@@@...............@@@@......@@@@@@@@@@@...@@@@@@@@@@@@@@@@@
.....@@@@@@@@@................@@@@........@@@@@@@@....@@@@@@@@@@@@@@@@@
........@@@@...................@@..........@@@@@........@@@@@@@..@.@@@.
&gt;&gt;&gt;&gt;^ ^&lt;&lt;&gt;&gt;^ ^&lt;&lt;&lt;&lt;&lt; &gt;&gt;&gt;^&lt;&lt;&lt; ^^ ^
&gt;,&lt; show the path of the detection algorithm
</PRE></TD></TR>
</TABLE>
<P>
The latest version of GOCR may use different algorithms.
You have to look at the sources learn more.
<H1><A NAME="SEC10">
Black/White, Gray and Colors</A>
</H1>
For simplicity colored images are converted to gray internally.
That means a red text on green background will not be detected.
You should use your own filter for this purpose.
<P>
If the original image is gray, a critical value is calculated to
extract characters from the background. This can fail, if images are
on the scanned page or tha scan is bad (dark edges or borders).
It is difficult to overcome this problem because graylevels are mostly
restricted to the 8 bit limit (16 bit would help to overcome this problem).
<P>
Black/White images are internally converted to gray with two levels (0 and
255).
<P>
The lowest 4 bits are not used, because they are used by internal functions
(this can be changed in future).
<P>
After calculation of the threshold value (otsu.c) the brightness of
every pixel is recalculated to a new internal threshold value
of 160 (128+32).
This is a bit above the middle of the 8 bit range. The idea is to
make the live easier for the other routines. Pixels which does not sure
belong to the white or black ones get a value near the threshold value.
Some routines can use this bit of more information to ignore outriders.
Second point is, that this is necessary for using lowest for bits
without destroying image informations.
<H1><A NAME="SEC11">Pictures on scanned pages</A>
</H1>
At first all objects on the scanned page are detected.
Objects are clusters of black pixels.
Pictures are detected if they are larger than 4 times the mean size of
all objects. This rule is very simple and can fail some times.
But it works fast and mostly the result is ok.
<H1><A NAME="SEC12">Tools</A></H1>
<P>
<DL COMPACT>
<DT>pbmclean:</DT>
<DD>This program is written by Angus Duggan and Jef Poskanzer.
It cleans up ``snow'' on bitmap images.
</DD>
<DT>pnmtools:</DT>
<DD>This tools are used to convert different image-formats to
easy readable PNM (PBM,PGM,PPM) format.
GOCR uses the popen-routine to call this programs if the
suffix of the filename matches to a list in pnm.c.
This will fail if pnmtools are missing.
</DD>
</DL>
<H1><A NAME="SEC12b">related projects (to learn from)</A></H1>
<P>
<DL COMPACT>
<DT>unpaper:</DT> <!-- Dec05 JS -->
<DD> <a href="http://unpaper.berlios.de/">unpaper</a> -
post-processing scanned and photocopied book pages,
written by Jens Gulden 2005, GPL
</DD>
</DL>
<H1><A NAME="SEC13">glossary</A> </H1> <DL COMPACT>
<DT>font series:</DT> <DD>bold, condensed</DD>
<DT>font shape: </DT> <DD>normal, italic, slanted, sc... </DD>
<DT>points:</DT>
<DD>length unit used for font size, 1/72 inch,
but I do not know its exact relation to the font size (height?
totalheight? width? 10pt and 300dpi results in 40 pixel heigh font?)
</DD>
<DT>sans serif:</DT>
<DD>font without the (often thin) lines on the ends
of the character
</DD>
<DT>descewing:</DT>
<DD>compensation of (slightly) rotated text
</DD>
</DL>
<H1><A NAME="SEC14"> More information?</A> </H1>
<DL COMPACT>
<DT>&middot;</DT>
<DD>see "/usr/share/doc/package/tetex/texmf/.../fntguide.dvi"
in the documentation of the tetex package
</DD>
<DT>&middot;</DT>
<DD>the fonts-HOWTO file is helpfully too
("www.faqs.org/faqs/fonts-faq/")
</DD>
<DT>RTF:</DT>
<DD> RichTextFormat - does someone have a good documontation?
</DD>
</DL>
<H1><A NAME="SEC15"> About this document</A> </H1>
This Document was originaly written in LaTeX.
In May 2002 Joerg has convertet it to HTML. The reason is, that
you can read it now directly and you does not need to have LaTeX and
Ghostscript installed on your computer to read it.
As a side effect you do not need tetex package to build the gocr.rpm-package.
A good viewer to read this document is lynx, links or w3m.
<BR>
<HR>
<ADDRESS> jNOschulen-at-gSmPAMx.de (remove NO+S+PAM) </ADDRESS>
</BODY>
</HTML>
<!---
%
% -----------------------------------------------------------------
% # v
% # v mark connected points via fifo-stack
% ### *<< and 3bit direction code per pixel,
% # # ^ ^ should be better for overlapping letters
% ### ^<<
% --------------------- point connections ----------------------
% searching next nearest point
% /----\
% | |
% ^ | ## |
% | | ## |
% | | | |
% | \--/ |
% \------/
%-------------------------------------------------------------
%
--->