I have been carrying out some experiments to recognise hand written characters from binary image forms through transfer learning from EMNIST data which is an extension of MNIST covering all Alphabets.

As we know MNIST data contains DN (digital number) in range 0-255, same is the case with EMNIST, so to perform transfer learning, I preferred to convert the EMNIST data into binary form. Results weren't that great, so I thought why not have a discussion on open forum.

1. How good is deep learning while dealing with binary images when compared to gray scale or continuous pixel images?
2. You might have heard about NIST special database from which the popular MNIST data is created. The original data consists of binary values but that was processed into gray scale by yann le cunn using anti-aliasing resampling algorithm. So if we have a set of binary image characters wouldn't it be wise to convert into same form like cunn's MNIST data and then use the model trained on MNIST for prediction rather than converting MNIST into binary and training a model on it.