search

qjadud1994 / CRNN-Keras

CRNN (CNN+RNN) for OCR using Keras / License Plate Recognition
MIT License
527 stars 191 forks source link

how do training ? #3

Closed PythonImageDeveloper closed 4 years ago

PythonImageDeveloper commented 6 years ago

Hi, Please give me more information about training and prepare dataset, how do apply labels for this ? please put this walk-through steps. thanks.

qjadud1994 commented 6 years ago

Sorry. I confirmed it a little late. I will update the training detail in the readme. Please check.

PythonImageDeveloper commented 6 years ago

Thanks for your updating. in your opinion, If we have photos that are not clean(include images that have border/ viewing angle/ .... ), that is not precisely the photo frame, then it works well? in your examples shows 5 types of plates that are very clean and no border of car and viewing angle. i mean such this example: https://www.google.com/search?q=number+plate+car+view&tbm=isch&source=iu&ictx=1&fir=Tr8CCfj7U6fm6M%253A%252CL6kIZpJJ4GdU4M%252C_&usg=AFrqEzcuLR8riujWQc-QAruDFkXn0j21qQ&sa=X&ved=2ahUKEwi0jp28ws3cAhUFDOwKHdinAU0Q9QEwB3oECAYQDg#imgrc=Tr8CCfj7U6fm6M:

my question 2: The most of plates have logo of country and name of country, in this case how is it work?

qjadud1994 commented 6 years ago

In my opinion, I think that there will be a lot of performance difference according to training data. If you have a lot of real license plate images, I think it will work well.

In addition, I know that scene text recognition models such as GRCNN, RARE, and FAN as well as CRNN works well in real environment.

PythonImageDeveloper commented 6 years ago

GRCNN is CNN + GRU?

qjadud1994 commented 6 years ago

No! GRCNN is Gated Recurrent Convolution Neural Network for OCR

It is similar to CRNN, but there is a difference in that certain parts of the CNN layer have been replaced by Gated Recurrent Convolution Layer.