Avoid text recognition in image areas of the page

[zuphilip]

In several tests ocropus/ocropy tries to recognize some text within an image area in my pages. How can this be avoided? Here is an example of such a page: normal-beispiel-mit-bild

I run the same sequence of commands as in run-test and ocropus/ocropy recognizes two columns one with the actual text and the other with some nonsensical symbols from the picture.

[tmbdev]

Generally, that requires good text/image segmentation, which isn't part of ocropy yet.

The plan is to provide a 2D LSTM implementation to learn text/image segmentation. For that that, I've been working on the CLSTM implementation.

[zuphilip]

Okay, I see. Can you say something about an ETA for such a feature? Days, weeks, months?

[tmbdev]

It's probably going to be months; it's a pretty hard problem to solve in general, and there is only limited training data.

Here is a paper describing an older approach: http://www.csse.uwa.edu.au/~shafait/papers/Bukhari-Text-Image-Segmentation-DAS10.pdf

[JamesOwers]

I'm doing my masters thesis on this topic at the moment and would like to get involved in implementing segmentation useful for ocropy. @tmbdev, are you looking for people to help?

[ankitag9]

@tmbdev @danvk I have been using ocropy for some time now and have made a few changes as well.. I am planning to work on ML based image segmentation next.. do you guys have some alpha version code of the same on which i can build?

[ChillarAnand]

Any updates on this issue?

[zuphilip]

Old C code is here https://github.com/michaelyin/ocropus-git/blob/ba6930627f3f81392a05a02bfcf7dac2595c35cf/ocr-layout/ocr-text-image-seg.cc and some Python code (~2012) https://github.com/zuphilip/ocropus-from-searchcode/blob/master/ocropy/ocropus-tiseg resp. https://github.com/zuphilip/ocropus-from-searchcode/blob/master/ocropy/OLD/ocropy/oldsimpleti.py

[amitdo]

Both Tesseract and ocropus 0.4 use the Leptonica library to do text/image segmentation.