About size of input images for training

qjadud1994 / CRNN-Keras

CRNN (CNN+RNN) for OCR using Keras / License Plate Recognition

MIT License

521 stars 191 forks source link

About size of input images for training #47

Closed Bao150297 closed 4 years ago

Bao150297 commented 4 years ago

I have two question:

The first, I have some real traffic images that I want to use for training model. But when I crop it to get only the plate, it maybe small ( ~ 50px). What size of the input images should be? And does it effect significantly to the accuracy of the model after training?
The second, I want to train model for recognize both two and one layer plate. So how many image of each class should be prepared for a accuracy model? Thanks in advance!

ZdsAlpha commented 4 years ago

Looking through code it looks like height of image is 64 pixels and you can have arbitrary width as it pads/crops the rest of image that does not fit.

qjadud1994 commented 4 years ago

Here, 128x64 is used. Customize the image size to suit you.
The number required depends on the characteristic of the training data. Also, due to the structure of the CRNN, two-line license plates may not be recognized well.