search

qjadud1994 / CRNN-Keras

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

very bad result in the real-plate images #24

Open ghost opened 5 years ago

ghost commented 5 years ago

Hi, I've trained the VGG-GRU network for around 700k artificially generated images, and i got 96% accuracy for test data (about 140k generated image), but when i test the trained the model on real-images , the results are very bad, why ?

qjadud1994 commented 5 years ago

Due to overfitting, it is no wonder that it does not work properly in real images.

so, I recommend pre-training with synthetic images and fine-tune with real data.

danishansari commented 5 years ago

@qjadud1994 How would you recommend fine-tuning, should we use differential learning rates, i.e. lower lr for CNN and higher for RNN? Or a global lower learning rate like 1e-3 or smaller? Or because it's using Adadelta, this is not required at all, and we can use default lr(1.0)? @mzeynali did you get it working?

PythonImageDeveloper commented 5 years ago

@qjadud1994, Because I used Early Stopping callback, in my opinion, Should not happen the overfitting problem, right?