Thank you for the suggestion. Will give tesseract 4.0 a try. I hear that tesseract 4.0 uses LSTM neural network, so its performance will be much better, especially for Chinese, but it may be much slower, is that true?
By the way, I have also tried tweaking the parameters of tesseract 3.05, and have significantly improved the results with the following parameters: assume_fixed_pitch_char_segment 1 textord_use_cjk_fp_model 1 textord_old_xheight 1 textord_min_xheight 60 textord_noise_hfract 0.1 On Thursday, September 21, 2017 at 4:01:26 AM UTC-7, shree wrote: > > You will have much better results if you use the new version of tesseract > from https://launchpad.net/~alex-p/+archive/ubuntu/tesseract-ocr > and the traineddata files from > https://github.com/tesseract-ocr/tessdata_best > > ShreeDevi > ____________________________________________________________ > भजन - कीर्तन - आरती @ http://bhajans.ramparivar.com > > On Thu, Sep 21, 2017 at 2:44 PM, wei ren <[email protected] <javascript:> > > wrote: > >> I am new to OCR and tesseract. Please forgive me if I ask some "stupid" >> questions. >> >> I try using tesseract 3.04.01 to recognize the Chinese characters in the >> attached two images and get absurd results, so I merge the two images into >> one and use the merged image yueyue.title.exp0.tif to train a new model. >> Below are the steps: >> >> 1. Create the box file. >> >> $ tesseract yueyue.title.exp0.tif yueyue.title.exp0 -l chi_sim >> batch.nochop makebox >> >> 2. Correct the errors in the box file in jTessBoxEditor. >> >> I fix the segmentation errors and assign the correct Chinese characters >> to the segmentations. >> >> 3. Train the new model. >> >> $ tesseract yueyue.title.exp0.tif yueyue.title.exp0 nobatch box.train >> $ unicharset_extractor yueyue.title.exp0.box >> >> 4. Define a font_properties file with the content. >> >> title 0 0 0 0 0 >> >> 5. Clustering. >> >> $ shapeclustering -F font_properties -U unicharset yueyue.title.exp0.tr >> $ mftraining -F font_properties -U unicharset -O unicharset >> yueyue.title.exp0.tr >> $ cntraining yueyue.title.exp0.tr >> >> 6. Prefix all the files with "title.". >> >> $ mv unicharset title.unicharset >> $ mv inttemp title.inttemp >> $ mv pffmtable title.pffmtable >> $ mv shapetable title.shapetable >> $ mv normproto title.normproto >> >> 7. Put all the files together. >> >> $ combine_tessdata title. >> >> 8. Copy the new model to the tesseract-ocr tessdata directory. >> >> $ sudo cp title.traineddata /usr/share/tesseract-ocr/tessdata/ >> >> Then I type the following command to recognize again the Chinese >> characters in the merged trained image. >> >> $ tesseract yueyue.title.exp0.tif stdout -l title >> >> Both the expected result is "老妇人和母鸡", but the actual result of the first >> page is "老 老老老妇 人老妇母老鸡老" and the actual result of the second page is >> "老老妇人和母老鸡". I generate a box file using the new model which is also >> attached, >> >> $ tesseract yueyue.title.exp0.tif yueyue.title.exp0 -l title batch.nochop >> makebox >> >> , and find that although tesseract only assigns the characters in the new >> model to the segmentations, it can't get the correct segmentations. As you >> can see, three characters are split into two segmentations, respectively. >> But when I correct the trained box file, I have merged those two >> segmentations into one. >> >> >> >> <https://lh3.googleusercontent.com/-r8UG3Svsbpo/WcN_98MjS7I/AAAAAAAAU8M/4ZMvHYfgOQ8OVp_fHIw__uZmTA6rFhyEgCLcBGAs/s1600/box2.png> >> >> <https://lh3.googleusercontent.com/-Wga1p7T579U/WcN_18CI2VI/AAAAAAAAU8I/Yvm9IB5zOGIXsbdcugPYTgiMRxbC02TTQCLcBGAs/s1600/box1.png> >> >> >> >> <https://lh3.googleusercontent.com/-Wga1p7T579U/WcN_18CI2VI/AAAAAAAAU8I/Yvm9IB5zOGIXsbdcugPYTgiMRxbC02TTQCLcBGAs/s1600/box1.png> >> >> >> >> <https://lh3.googleusercontent.com/-Wga1p7T579U/WcN_18CI2VI/AAAAAAAAU8I/Yvm9IB5zOGIXsbdcugPYTgiMRxbC02TTQCLcBGAs/s1600/box1.png> >> >> >> >> <https://lh3.googleusercontent.com/-Wga1p7T579U/WcN_18CI2VI/AAAAAAAAU8I/Yvm9IB5zOGIXsbdcugPYTgiMRxbC02TTQCLcBGAs/s1600/box1.png> >> >> >> >> <https://lh3.googleusercontent.com/-Wga1p7T579U/WcN_18CI2VI/AAAAAAAAU8I/Yvm9IB5zOGIXsbdcugPYTgiMRxbC02TTQCLcBGAs/s1600/box1.png> >> >> I have tried specified the font as bold and/or fixed in font_properties >> and it doesn't help. I have also tried various page segmentation methods >> and it doesn't help either. >> >> >> I also attach the trained tessdata here so you can easily reproduce the >> problems. Any hint or suggestion will be highly appreciated. >> >> <https://lh3.googleusercontent.com/-Wga1p7T579U/WcN_18CI2VI/AAAAAAAAU8I/Yvm9IB5zOGIXsbdcugPYTgiMRxbC02TTQCLcBGAs/s1600/box1.png> >> >> -- >> 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] <javascript:>. >> To post to this group, send email to [email protected] >> <javascript:>. >> 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/18590868-ba1e-457d-8953-b002987d497d%40googlegroups.com >> >> <https://groups.google.com/d/msgid/tesseract-ocr/18590868-ba1e-457d-8953-b002987d497d%40googlegroups.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/4a702893-da3f-4b26-998e-aba4f04271cb%40googlegroups.com. For more options, visit https://groups.google.com/d/optout.
