Is there any algo that generates more robust datasets from inital sets? or block the system from learning from a set a features?

During my presentation Machine Learning Empowered by Python, I created a real-time digit recognition demo. The training samples were black and white. I was looking for a way to train my system independently of the pixel colors so it could work on opposite contrast or any color (i.e.: color independent). In order to do so, I have forced a preprocessing to black-and-white, but I am pretty sure there is a better way that requires less manual intervention.

There might be a way to add constraints to the training so it doesn't train from pixel color.