Machine Learning and Friends Lunch
Leveraging Character Similarity to Improve Recognition of Novel Fonts
Jerod Weinman
UMass
Abstract
By unifying language and contextual information, people can
"read" text in
fonts that bear no resemblance to our usual characters. In this
talk, I will
describe a probabilistic model for character recognition
inspired by
cryptogram solving. Our unified framework obviates ad hoc
iterative methods
and eliminates piecewise processing that can result in
unrecoverable errors.
Using learned, soft constraints that encourage characters of
similar
appearance to have the same label, we can improve the ability of
our system
to read signs around downtown Amherst.