search

xinntao / Real-ESRGAN

Real-ESRGAN aims at developing Practical Algorithms for General Image/Video Restoration.
BSD 3-Clause "New" or "Revised" License
27.57k stars 3.46k forks source link

will it work on text images? #78

Open YC7225 opened 2 years ago

YC7225 commented 2 years ago

Do you guys think will it work on text images like Driving License images capture from phone with blur and noise? will it helpful for remove all this things and perform OCR on it?

xinntao commented 2 years ago

It will work on text images. But you need to finetune on those images with texts.

YC7225 commented 2 years ago

what are the fine tuning parameter required for that?

xinntao commented 2 years ago
  1. Data is important. So we need to collect such text images for fine-tuning the pre-trained models
  2. For better performance, we also need to tune the hyper-parameters for synthesizing low-quality images
YC7225 commented 2 years ago

okay! so I have tried on some text images so far text is small and lots of text in image text style or font style will change that will cause the ocr problem..so I am thinking to train on this dataset what do you think based on your knowledge for getting better result and at the time of generation we will get better result.. link of the dataset:- https://github.com/JasonBoy1/TextZoom

R2Bb1T commented 1 year ago

okay! so I have tried on some text images so far text is small and lots of text in image text style or font style will change that will cause the ocr problem..so I am thinking to train on this dataset what do you think based on your knowledge for getting better result and at the time of generation we will get better result.. link of the dataset:- https://github.com/JasonBoy1/TextZoom

I'm doing exactly the same thing recently, did you work it out? I've tried those two way mentioned by @xinntao.

R2Bb1T commented 1 year ago

@YC7225

javierquin commented 3 months ago

Did you have any luck? I need to restore text from blurry/low quality images