[Bug-ocrad] Technical documentation summary readme.txt, page skew, Hough transform.

[Chris K. Skinner]

This will be long, so please be patient and see if you can read all of this 
...

I'm interested in various aspects of computer science including image 
processing, neural networks,
expert systems, semantic analysis.  I've got books on such topics.

As you are probably aware, there are software patents in some countries.

If you had some kind of outline of the algorithms that were applied per 
version of the software that would greatly help someone new coming in fresh 
off the street to gain a quicker understanding of stuff in general, and 
probably demonstrate to the world at large that you have invented something 
new that could not be patented / stolen / claimed by some greedy corporate 
dudes.

I have just downloaded and tried to compile your source, but it failed.  Now 
to understand your work, I have to look into each source to try to realize 
what is being done now.  To see what was done in previous versions that 
either did not work as intended or was tried and abandoned, I would have to 
repeat this analysis on your older version sources, and compare it to the 
current version.

Do you have any design notes, bibliographic citations, web links to 
information that you've made use of , release notes for what algorithms are 
being used / abandoned.

I've been using the following OCR software since about 1993:
WinFax, Calara WordScan Plus, Caere OmniPage.

> From my experience with these, the amount of OCR errors goes way up if the 
page skew (orientation, angle, rotation, rotate) is not exactly aligned to 
zero degrees.  When the angle is off, then the bounding boxes around each 
page element, each column, each line of text, each character is wrongly 
positioned to create huge amounts of recognition errors.  Consider that when 
a high resolution scan is done that recognition should probably improve 
because the information is rich with nice amounts of redundant information 
clues as to what is present on the page.

But the long horizontal lines of text then become very long "sets of stripes 
of pixels."  With such long stripes, it is more likely that instead of there 
being a one or two pixel error from page skew, it can be much higher.  If 
the recognition algorithms do not account for this, and instead determine 
bounding box regions for recognition too early and presume a zero page skew 
angle error, the results shall be/are very bad.

In the J. R. Parker book w/CD ROM "Algorithms For Image Processing And 
Computer Vision" that I have read, the author provides a couple of algorithm 
suggestions for combating the page skew angle issue.  A Hough-transform when 
applied to the dots of the bottoms of the bounding boxes of glyphs results 
in a page skew angle in degrees (with his source code, that is).  By 
applying an image rotation that eliminates the skew, better recognition 
shall result.  (Unfortunately, he does not, however, present the source code 
for determining the bounding boxes of glyphs so that it is not easy to 
demonstrate that this algorithm will work especially on larger regions of 
text.)

Another approach is to use angle-independent Complex-Number-Coefficient 
Neural Networks to use as feature recognizers.  The Japanese promoter of 
these neural networks says that they are Affine-Transform insensitive, and 
thereby can recognize a pattern that has been so transformed.
http://mathworld.wolfram.com/AffineTransformation.html
http://www.google.ca/search?num=20&hl=en&newwindow=1&safe=off&q=Affine
" 
http://www.google.ca/search?num=20&hl=en&newwindow=1&safe=off&q=Affine+Complex-Number+Coefficient+Neural+Networks 
"
This too is just a theory.  I don't have a copy of any books on 
Complex-Number-Coefficient Neural Networks, or any source code from a 
competent mathematician who has converted the advanced mathematics into 
working C++ code examples.  Often these theoreticians are not interested in 
the practical applications of their work and are more interested in the 
expressions of their ideas as continuous functions expressed as 
N-dimensional differential equations (or something much less understandable 
to me anyway).

Thanks for any help that you could provide me in helping understand your 
project so that I might possibly provide you with suggestions for 
improvements.

Kindest regards, C.