That is going to depend on what you want the features to measure.
No-one knows what the best features are for any particular kind of classification or neural networks or deep learning. No one. There are just features that have been useful in the past on particular input images, features that were kept after Feature Reduction on that training set, and that set of input features.
