For downloading new validate code images, you can execute the program download_images.py.
It will download validate code images from the NTNU course taking website by sending GET requests to https://cos1s.ntnu.edu.tw/AasEnrollStudent/RandImage.
To change how many validate code images to download, just edit the variable in download_images.py.
After download validate code images from the website, you will need to label them by yourself.
For this task, I wrote a simple program labeling.py.
These are 600 validate code images & labels I used.
Feel free to download and use them.
The architecture of the best model I found is in the file best_model.py.
To use it, you can try understanding how the program predict.py works, and do some adjustments to fit your need.
However, you will need the weights file (val_loss.h5) that I have trained, which is at here.
Remember to edit the weights file (val_loss.h5) path in the first line of best_model.py to where you put the pretrained weights file.