Hi, I'm new to tesseract too, and also working on the fine-tuning . Wish this could do any help to you.
=> After lstmtraining finished, I got only* "basetrain.log"* and > * "base_checkpoint"* but I cannot find the final "khm.traineddata". and I > really don't know why. Could you please help me? > You can use the stop_training flag as following to create traineddata lstmtraining --stop_training --continue_from trainhalfwidth/jpnhw_checkpoint --traineddata tessdata/jpn.traineddata --model_output trainhalfwidth/jpnhw.traineddata Besides, I suggest you can firstly try the "Fine Tuning for Impact", the "from the scratch" seems did not works for me, cause the langdata provided is for tesseract 3 training. ps. Sorry for my bad english. 在 2017年11月21日星期二 UTC+8下午7:47:14，phyrum sk写道： > > In Khmer language, there are 2 types of fonts. > 1. Pre-Unicode (Limon) - 19 fonts such as Limon F1, Limon F2 ...etc and > 2. Unicode fonts such as Khmer OS, Khmer OS Battambong ... etc > > I tested using latest tesseract tessdata i.e khm.traineddata_fast and > khm.traineddata_best (khm.traineddata_fast gaves better accuracy) and > measured its accuracy using ISRI tools. > The average of Khmer Unicode Character and Cluster accuracy are more than > 80% but for Pre-Unicode (Limon) is around 60%. > > + Most of the Khmer old (Law) documents [images,pdf] files were produced > using Khmer Legacy fonts (Limon). so our goal is to Fine Tuning the > existing Tesseract khm.traineddata with Khmer Pre-Unicode fonts. So the OCR > engine can recognize those image files and produce output as unicode text > with better accuracy. > > A. Problem 1: Since I am naive to Tesseract and LSTM-Neural Network, I > want to understand how Tesseract 4.0 LSTMtraining work so I follow the > tesseract wiki tutorial and tried to trained lstm from the scratch using > Khmer OS Font and tesseract's langdata for khmer. > > Here are the command I used: > > training/tesstrain.sh --fonts_dir /usr/share/fonts/truetype/khmeros-ttf > --fontlist "Khmer OS" --langdata_dir langdata --lang khm --linedata_only > --noextract_font_properties --tessdata_dir /home/phyrum/tesseract/tessdata > --output_dir khmtrain > > training/lstmtraining --debug_interval -1 \ > --traineddata khmtrain/khm/khm.traineddata \ > --net_spec '[1,36,0,1 Ct3,3,16 Mp3,3 Lfys48 Lfx96 Lrx96 Lfx256 O1c111]' \ > --model_output khmoutput/base --learning_rate 20e-4 \ > --train_listfile khmtrain/khm.training_files.txt \ > --max_iterations 10000 &> khmoutput/basetrain.log > > => After lstmtraining finished, I got only* "basetrain.log"* and* > "base_checkpoint"* but I cannot find the final "khm.traineddata". and I > really don't know why. Could you please help me? > > Note I used Ubuntu 16.04 LTS and My Tesseract Version is > 4.00.00dev-691-gfb359fc > > B. Problem 2: In Pre-Unicode (Limon) fonts, each character is represented > using Latin-based codes, not in unicode. > Process: > - I converted langdata/khm/khm.training_text and other files in langdata > into Limon > > - Run command *training/tesstrain.sh* and *training/lstmtraining* (the > same as above command except parameter --fontlist and --fonts_dir is > changed) > > - The output of .box file and unicharset were in latin base so It cannot > recognized output text as unicode text. > > Is it the correct way to do it? Could you please give me some advices? or > Is there any where in the Tesseract source code that should be modified so > I can train by using input text as Limon format but the OCR recognized > output is in khmer unicode text. > > Thanks with best regards, > Phyrum > > > -- 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/8080fe5c-69df-4cf6-80e6-4a948516bafd%40googlegroups.com. For more options, visit https://groups.google.com/d/optout.
