https://code.google.com/p/tesseract-ocr/issues/detail?id=1351
What steps will reproduce the problem?
1.Use tesseract build with OpenCL.
2.Pass full color image with width which is not multiple of 32.
3.Recognition is way too slow and does not recognize anything.
I read the article on http://www.sk-spell.sk.cx/tesseract-meets-the-opencl-first-test and decided to give OCL a try. The initial result was as per point 3 above. After some debugging I figured the problem is that the OCL version of threshold rect generation does not account for padding bits in the output pix lines. To prove my discovery I made a quick fix in oclkernels.h replacing the definition of kernel_ThresholdRectToPix
Just a reminder: it is necessary to force OCL kernel recompilation after changing this source (e.g. delete “kernel - <device>.bin” from the exec folder).
The fix is working but I am not sure about it since the original source apparently works for other people (as per the article). If I am right the OS/GPU are irrelevant since the bug is algorithmic, but mine are Windows/AMD. Also similar fix is applicable to kernel_ThresholdRectToPix_OneChan(), but there the input array might have some padding bytes as well, so its indexing will need further adjustments. I can come with some prove/fix for it either - I have not played with it yet.
Disclaimer: I have no prior experience with image processing and tesseract source or with GPU computing and OpenCL (but please do explain if I am wrong).
to improve correctness and compatibility with
external programs, particularly ghostscript.
We will start mapping everything to a single glyph,
rather than allowing characters to run off the end
of the font.
A more detailed design discussion is embedded into
pdfrenderer.cpp comments. The font, source code
that produces the font, and the design comments
were contributed by Ken Sharp from Artifex Software.
Font recognition was poor, due to forcing a 1st and 2nd choice at
a character level, when the total score for the correct font is often
correct at the word level, so allowed the propagation of a full set
of fonts and scores to the word recognizer, which can now decide word
level fonts using the scores instead of simple votes.
Change precipitated a cleanup of output data structures for classifier
results, eliminating ScoredClass and INT_RESULT_STRUCT, with a few
extra elements going in UnicharRating, and using that wherever possible.
That added the extra complexity of 1-rating due to a flip between 0 is
good and 0 is bad for the internal classifier scores before they are
converted to rating and certainty.
Tha, Vie, Kan, Tel etc.
There is a new overlap detector that detects when diacritics
cause a big increase in textline overlap. In such cases, diacritics from
overlap regions are kept separate from layout analysis completely, allowing
textline formation to happen without them. The diacritics are then assigned
to 0, 1 or 2 close words at the end of layout analysis, using and modifying
an old noise detection data path.
The stored diacritics are used or not during recognition according to the
character classifier's liking for them.
Eliminated the flexfx scheme for calling global feature extractor functions
through an array of function pointers.
Deleted dead code I found as a by-product.
This CL does not change BlobToTrainingSample or ExtractFeatures to be full
members of Classify (the eventual goal) as that would make it even bigger,
since there are a lot of callers to these functions.
When ExtractFeatures and BlobToTrainingSample are members of Classify they
will be able to access control parameters in Classify, which will greatly
simplify developing variations to the feature extraction process.
