Hi,
There are generally 2 ways to deal with such problems. Either you rectify the orientation of image and train the model for straight up digits or you can add images with different orientation in your training data and let your deep model handle such complexity. I would prefer second approach because correcting the orientation is difficult task and cannot be generalised effectively which will result in error and in turn affect the detection also. So I would suggest you to augment your data to take care of such random rotation. it will make your model more robust. The easy way to deal with this is to use imageDataAugmentor function during training. It will help you create digits image with random rotation and then you can train your model with such variety, in the end you will obtain a robust model which can take of rotation by itself.
I hope this helps.
Cheers.
