I am new to deep learning and I am trying to train a NN to recognize house numbers gathered from street view. I have already managed to recognized MNIST sequence of hand written digits by means of a CNN. In that case I have trained the CNN with a thousands of examples consisting of 1 from up to 5 digits sequences created by me.
I have thus thought to do the same thing for SVHN, I have cropped the numbers and I now have a large dataset of single digits from street view. However the MNIST data set is very homogenous (i.e. images have same size, same angle, same color, etc.) while SVHN is obviously not. Trying to create sequences of digits from different SV photos led to weird examples that are actually not representative of the real word objects I want to model.
So I am wondering if it was better to train a NN on the single digits and then use it to recognize sequences. Would it be possible? How?