Resolution Learning in Deep Convolutional Networks Using Scale-Space Theory