search

HsiehYiChia / Scene-text-recognition

Scene text detection and recognition based on Extremal Region(ER)
MIT License
163 stars 55 forks source link

Now i am getting this error after running the .exe file #2

Closed arsalan993 closed 4 years ago

arsalan993 commented 6 years ago

.\canny_text.exe -icdar Error: the input file is not opened!! Error: the input file is not opened!! Error: the Transition Probability Table file is not opened!!

Btw i also tried to open up an image but same error

HsiehYiChia commented 6 years ago

Try to put the dictionay/ or dictionary/tp_table.txt in the same directory with .exe. Sorry for my terrible readme file and SW dependency, I should make sure it is able to run when you clone it. In fact, I stopped to update the project and my environment is not available now, therefore I can't update it now. Maybe I'll in the future when I got spare time, I mean maybe :P

arsalan993 commented 6 years ago

So i ended up copying all data from /x64/release folder to main directory and it worked like charm

arsalan993 commented 6 years ago

Keep in touch i might be needing more help ,

arsalan993 commented 6 years ago

Is this algorithm heavily dependent on Font type.. how can i add more fonts.. does it requires training or not. How much is the accuracy and on which relevant font size in an image it works fine

HsiehYiChia commented 6 years ago
  1. Yes, the OCR model is based on edge feature, and therefore depends on font.
  2. To add font, you need to put every single letter in ocr_classifier/FONTNAME/, arrange them by font face and case. You can refer to ocr_classifier/Calibri/Normal to see the file layout. Then add font name to font_name in ocr_classifier/get_ocr_data.py. The script will do all the jobs and train model by libsvm, so make sure to put libsvm in C:\. Unfortunately, the script also need a .exe which I didn't not commit it and has lost, so I don't think you can add new font.......
  3. The accuracy for end-to-end detection on ICDAR 2013 is roughly 70% with fine tune. As for end-to-end recognition, I didn't test for it, but ICDAR 2013 winner is roughly 40%.
  4. This algorithm works best with san-serif font.
  5. Currently, Deep Learning is dominating this field, end-to-end recognition accuracy is higher than 90%, maybe you should try it. Click me to see ICDAR 2017 result