If you binarize an image you loose of course the edge steepness details. Usually one is interested to generate the same binarisation behaviour over the whole image which is tricky because of different illumination effects of real-world images. A well known standard method (old, but effective) is Otsu's approach. Here you will find some input regarding binarisation: (http://homepages.inf.ed.ac.uk/rbf/CVonline/LOCAL_COPIES/MORSE/threshold.pdf)