The way you know how to train with boxes seems very much like version 3.0.5 learning. Im currently only interested in LSTM learning (aka Tesseract 4.*). Since i last wrote i managed to dig up something that actually helped me a lot in both understanding the process and actually training data. So I was pointed at the following github repo <https://github.com/OCR-D/ocrd-train>. I run my training on Win10 64bit with cygwin. I just had to put the files from repo in folder and inside execute (order counts) make leptonica and then make tesseract, with these steps you compiled tesseract and leptonica. (other links: 1 <https://groups.google.com/forum/#!topic/tesseract-ocr/CLFM6dw5HIM>, 2 <https://groups.google.com/forum/#!searchin/tesseract-ocr/l.bolzani%7Csort:date/tesseract-ocr/be4-rjvY2tQ/32evtMHlAQAJ>, 3 <https://groups.google.com/forum/#!msg/tesseract-ocr/COJ4IjcrL6s/C1OeE9bWBgAJ>) Before you start training there are some changes to be made on makefile and generate_line_box.py. I opened them with notepad++. In makefile you have to change variables according to your stuff and in .py you have to change a few lines (here <https://github.com/OCR-D/ocrd-train/issues/18> you can see why). If you make these changes you can run make training and it will get your tiff/text file pairs from ground-truth and auto create every training file. Essentially if you understand what the makefile does you will understand the process of training and the key to make it for your own specifications is to edit that makefile to your needs. What i mean for example its hardcoded for max 10000 iterations so if you want something else you have to change it in the script. The script itself executes codes that you can find on wiki but for me the script gave more details because i could see what steps are there and what for do i need them.
2019. február 5., kedd 15:32:05 UTC+1 időpontban WidmoPL a következőt írta: > > Hello, > Thanks for quick reply. > Yeah, from what i found, or rather didn't find, there is very little > documentation or it is poorly written,especially for version 4. > I only know how to train using boxes and then editing them in some > ''editor'' when you must check letters with rectangulares. This is I know > well by now,so if you need some info to this part I will gladly help. But > this method is very time expensive, despite results looks promising. Also > noticed that when you recognize text having trained data done this way its > most time better not to use standard trained data + your trained data, just > only use your trained data. > Example: > tesseract qwe.arial.exp0.png qwe.arial.exp0 -l xyz -> where > xyz is name of your trained data file > > AND NOT -> tesseract qwe.arial.exp0.png qwe.arial.exp0 -l eng+xyz > > > Thats all interesting i can say for now. > Thanks for reply! > > > W dniu poniedziałek, 4 lutego 2019 14:33:09 UTC+1 użytkownik Kristóf > Horváth napisał: >> >> Oh boy, where to start! So first of all you are not alone not finding any >> information. Currently i am a week ahead of you, so im gonna share what i >> found out. >> Lets start with training_files.txt. Whats inside? >> /home/kh/tesstutorial/engtrain/eng.Arial.exp0.lstmf/home/kh/tesstutorial/engtrain/eng.Impact_Condensed.exp0.lstmf >> What is it for? it holds the location of .lstmf files, which are required >> for training (tesseract 4.*) >> How do i get it? if you manage to create starter training files it will >> create it with it so you dont have to write it on your own. >> Whats the difference between eval and train? as far as i know the only >> difference is their place in the training proccess. So i think you compare >> results from train training_files to eval training_files. (Sadly im still >> stuck on this question) >> >> Thats all i know about these training files >> >> And for your second question: If thats possible i wanna know how. >> Currently for me it is a controversal question because there is no >> documentation and what you can find says that its not supported but others >> say oh you can do it ( >> https://groups.google.com/forum/#!topic/tesseract-ocr/0r8QvV3j8ew) >> >> Hopefully i could help, currently im working on a newbie friendly >> documentation for tesseract 4.* training so if you get some info on topic, >> i would appreciate if you would send it. >> Good luck for going further into topic. >> >> 2019. február 4., hétfő 12:15:13 UTC+1 időpontban Krzysztof Kanafa a >> következőt írta: >>> >>> Helllo >>> >>> I'm completely new in tesseract, first version I'm using is 4.0.0. Sorry >>> for noob question, but I really didn't find answer despite quite long >>> searching. >>> Its about these two options -> eval_listfile and train_listfile. What >>> exactly should be in these files? >>> Is there in train_listfile.txt should be list of images that we want to >>> be learned? for example images with scanned letters: >>> 1. first image "image1.tif" -> ABCDEF >>> 2. second "image2.tif" -> GHIJ >>> And in eval_listfile.txt a list of same named as images text files that >>> each contain good solution? >>> Example: >>> 1.first txt file "image1.txt" with text inside -> ABCDEF >>> 2. second txt file "image2.txt" ->GHIJ >>> >>> Or is it about completely something different? >>> >>> >>> >>> And second question ..: >>> when i ve already described examples... let's say i have such two TIF >>> files as mentioned before (photos with some serial numbers from a kind of >>> labels - each photo has one serial no). >>> And in two txt files I wrote by hand these serial numbers (just read it >>> from photo - one serial in one txt file). >>> >>> *Is there a way to just point these files *saying tesseract -> >>> >>> "here are some photo and here are some txt files telling what is on each >>> photo" >>> >>> and make him learn such way? for example if i have 60 of such tif+txt >>> paris, in this case selecting one by one letter in each tif (using BOX >>> files of course) will take forever, and what if i would have 50 pairs but >>> each tiff have 50 pages, that would take forever.. >>> Again sorry foor noob question, I read about tesseract for few days and >>> it all mixed up in head. >>> >>> Greetings >>> Krzysiek >>> >> -- You received this message because you are subscribed to the Google Groups "tesseract-ocr" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To post to this group, send email to [email protected]. Visit this group at https://groups.google.com/group/tesseract-ocr. To view this discussion on the web visit https://groups.google.com/d/msgid/tesseract-ocr/dcb382c3-ea0a-4aa1-aa68-145766bf4feb%40googlegroups.com. For more options, visit https://groups.google.com/d/optout.
