What makes it possible to search over font characters but not, eg, convolution tiles? Photo identification is done with a limited search space of “tiles” that

What makes it possible to search over font characters but not, eg, convolution tiles? Photo identification is done with a limited search space of “tiles” that the image is decomposed into, for convolutional NNs.