On Wed, Mar 20, 2019 at 9:57 AM 易鑫 <[email protected]> wrote:
> Thank you very much for your reply, your result is pretty good. > > You are right, I want to limit my unicharset. > I want to ask you a few questions: > > 1.What pre-processing have you done? only Binarisation,Rotation and > Deskewing? > I used irfanview interactively. Rotated to straighten the lines, converted to 2 color image and changed dpi to 300. I didn't test with oiginal image. Tesseract also does binarization. > > 2.From your result,chi_sim_tuned.txt, also contains some characters that > do not in the train_text file,such as "二"，“》:”，why? > I don't know. Probably they are there in the tessdata_best model and don't get fully overwritten in finetuning. > > 3. How to the choose the "max_iterations" value, I usually choose a large > number for the first time such as 10000 to let the model under overfitting > condition, then reduce the value gradually,make sure the model is good > finally. > Is there any good method to choose max_iterations? > Ray's recommendations for finetuning for font is 400 iterations. For plus-minus tuning to add a character is 3600. You should check an eval set (different from training set) around these numbers to find the minimum. > > > > > > > > > > > > > > Shree Devi Kumar <[email protected]> 于2019年3月20日周三 上午11:18写道： > >> >> ~/tesseract/src/training/tesstrain.sh \ >> --fonts_dir ~/.fonts \ >> --training_text ~/langdata/chi_sim/chi_sim_tuned.txt \ >> --langdata_dir ~/langdata \ >> --tessdata_dir ~/tessdata \ >> --lang chi_sim --linedata_only \ >> --noextract_font_properties \ >> --exposures "0" \ >> --workspace_dir ~/tmp \ >> --save_box_tiff \ >> --fontlist \ >> "NSimSun" \ >> "Arial Unicode MS" \ >> "SimSun" \ >> "Merchant Copy" \ >> "Merchant Copy Doublesize" \ >> "Noto Sans CJK SC" \ >> "Noto Sans Mono CJK SC" \ >> --output_dir ~/tesstutorial/chi_sim_trainnew >> >> >> mkdir -p ~/tesstutorial/chi_sim_tuned_from_chi_sim >> >> combine_tessdata -e ~/tessdata_best/chi_sim.traineddata >> ~/tesstutorial/chi_sim_tuned_from_chi_sim/chi_sim.lstm >> >> ~/tesseract/bin/src/training/lstmtraining \ >> --model_output ~/tesstutorial/chi_sim_tuned_from_chi_sim/chi_sim_tuned \ >> --continue_from ~/tesstutorial/chi_sim_tuned_from_chi_sim/chi_sim.lstm \ >> --traineddata ~/tesstutorial/chi_sim_train/chi_sim/chi_sim.traineddata \ >> --old_traineddata ~/tessdata_best/chi_sim.traineddata \ >> --train_listfile ~/tesstutorial/chi_sim_train/chi_sim.training_files.txt \ >> --debug_interval -1 \ >> --max_iterations 3600 >> >> ~/tesseract/bin/src/training/lstmtraining \ >> --stop_training \ >> --continue_from >> ~/tesstutorial/chi_sim_tuned_from_chi_sim/chi_sim_tuned_checkpoint \ >> --traineddata ~/tesstutorial/chi_sim_train/chi_sim/chi_sim.traineddata \ >> --model_output ~/tessdata_best/chi_sim_tuned.traineddata >> >> >> On Wed, Mar 20, 2019 at 8:46 AM Shree Devi Kumar <[email protected]> >> wrote: >> >>> Also, 10000 iterations for finetuning will lead to overfitting. >>> >>> I tried by using fewer fonts and adding a couple of English only fonts >>> that match the typeface of the image you shared. The output is improved >>> compared to tessdata_best. I assume that you want to limit your unicharset >>> based on your training_text (numbers, some English letters and some >>> Simplified Chinese characters). The image was pre-processed to B&W and >>> deskewed. >>> >>> I found that --psm 6 gives worse results both for tessdata_best and >>> finetuned, but the default psm gives better accuracy though there are >>> multiple blank lines for extra columns identified in --psm 3. >>> >>> See attached: >>> >>> >>> >> >> -- >> >> ____________________________________________________________ >> भजन - कीर्तन - आरती @ http://bhajans.ramparivar.com >> >> -- >> 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/CAG2NduUeONc98a%3DMiGE1Y1PGKK-Jb5vinDTPnEF%2BMvPUkT0nmw%40mail.gmail.com >> <https://groups.google.com/d/msgid/tesseract-ocr/CAG2NduUeONc98a%3DMiGE1Y1PGKK-Jb5vinDTPnEF%2BMvPUkT0nmw%40mail.gmail.com?utm_medium=email&utm_source=footer> >> . >> For more options, visit https://groups.google.com/d/optout. >> > -- > 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/CAPiKE21ywZpg%2BRtGj2BK9XxV87ivycnhp8nvaGSguaD%3DtKUN7w%40mail.gmail.com > <https://groups.google.com/d/msgid/tesseract-ocr/CAPiKE21ywZpg%2BRtGj2BK9XxV87ivycnhp8nvaGSguaD%3DtKUN7w%40mail.gmail.com?utm_medium=email&utm_source=footer> > . > For more options, visit https://groups.google.com/d/optout. > -- ____________________________________________________________ भजन - कीर्तन - आरती @ http://bhajans.ramparivar.com -- 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/CAG2NduWs5cctkn0OSF9UE2Fhhq7wsyE8xmFwwdj%2BAQVXfqNfFA%40mail.gmail.com. For more options, visit https://groups.google.com/d/optout.
