Remove old code which was used for Ocropus

[stweil]

Signed-off-by: Stefan Weil [email protected]

[stweil]

This should fix issue #2926.

[zdenop]

thanks

[amitdo]

SetFillLatticeFunc() was not part of the ocropus cut-line.

[amitdo]

fill_lattice_

[stweil]

@amitdo, maybe the code which refers to fill_lattice_ could be removed, too. Or is there someone who uses that?

[amitdo]

Maybe that code is/was used by Google internal version.

[amitdo]

Here's another SetSomethingFunc

tesseract/src/api/baseapi.cpp

Line 1993 in 23be532

void TessBaseAPI::SetProbabilityInContextFunc(ProbabilityInContextFunc f) {

probability_in_context_

https://groups.google.com/forum/?hl=en#!topic/tesseract-dev/YsJogVVFbGk

Ray 9/9/14

There is no code in tesseract to implement or train an n-gram model, only to make use of one.

There is a "hook" function and code to set it to an external function.
TessBaseAPI::SetProbabilityInContextFunc is used to set it.

We experimented with it in Google with some existing Google-specific n-gram library. It didn't help very much in most languages, although it does in cases where the dictionary is particularly weak for some reason, like CJK.

A single function is required that returns the probability of the current "character" (in general a utf-8 string) given the context of a previous utf-8 string.
It is fairly simple to write and train, given appropriate training data.

The difficulty is with balancing the strength of the n-gram model and the shape classifier confidences. For speech recognition, where typical system accuracies are 80-90%, it is easy to use language models to combine with a poor accuracy phoneme classifier, but in OCR, where typical word accuracies exceed 95% (for alphabetic, word-segmented languages) language models of any kind can easily start replacing infrequent, but correct results with more frequent, but wrong results, and nobody has properly solved the problem in a way that is mathematically and statistically rigorous. Looking at OCR output, it is easy to spot cases where language models could help, but not so easy to see the already-correct results that would be made worse.

[amitdo]

@stweil,

GetPageRes() is used in unittest/baseapi_test.cc

[stweil]

Thank you. Sorry that I missed that. Pull request #2962 should fix it.