|
We
present new techniques and results for word recognition and for
writer identification from handwritten documents. These problems
are complementary and require the development of features that
depend on word content but are independent of stylistic variations
for the first problem and depend on style but are independent
of content for the second. Accordingly, we have developed both
a set of style-dependent features and a set of style-independent
features. These features have been applied to two handwritten
document databases: 1) a signature database collected from company
timecards for word recognition and 2) a Chinese document database.
Results are presented for both. For word recognition, we present
a new metric for invariant image recognition. This metric can
be used to accurately recognize objects in spite of the presence
of noise, distortions in the shape of the image, and clutter which
overlaps and obscures parts of the object. For writer identification,
we present a novel method of unsupervised clustering of documents
based on the handwriting style of their authors.
|
