After !nding candidate CCs, we !lter them based on
various heuristics. Since most text has low variation in
stroke width, we !lter CCs whose stroke width standard
deviation is greather than 1/2 the mean stroke width.
We also !lter CCs with high variation in lightness and
color in the Lab space. Since most classi!able text in
images is not very small, we also !lter CCs that are less
than 10 pixels in width or height. Finally, we !lter CCs
with an aspect ratio greater than 10. The input image
masked by the !ltered CCs is shown in Figure 2g.