Updated CNN .keras model with input layer 256*256 RGB, and 1 output node ranging from 0 (no rotation) to 1 (180 degrees).
Its accuracy on the dataset in the link below is 99.51%, and the AUC (Area Under Curve) metric is 0.9990.
https://drive.google.com/file/d/1WENDw-Ggwy32KIdMlMmuF6NcO0gH3mC8/view?usp=sharingIt only recognizes 0 vs 180 degrees as there's another robust way to distinguish +- 90 from 0 and 180 degrees (another model under UTSC DSU).
Additionally, the dataset only contains upright images because the training program generates the 180-degrees counterparts on the fly.
As you will see, the model is only trained on text slices resulted from my proprietary preprocessing methods
(old version already on GitHub
https://github.com/digitalutsc/ml_image_tool, but I will update it soon),
So directly feeding the page images to the model may not work well.
I used MobileNetV2 with imagenet weights as the backbone, and the accuracy proves the method to be very successful.
