The goal of this project was to test the
use of deformable models for handwritten
digit recognition. To test the approach, we
collected a database of 440 handwritten
digits (0, 1, ..., 9) from 44 different
individuals, some of which are shown in
Fig. 1. Each letter was scanned and resized
to 32 x 32 pixel size.

Figure 1: Examples of handwritten digits from 4 different individuals are shown for the 10 classes.

For each digit we developed a
deformable model whose shape is deformed
to match the handwritten digit. While the
models were in vector form, we rasterized
them to 32 x 32 pixel size, so that they
could be compared to the given handwritten
digit. The idea was to test all the
deformable models against a given
handwritten digit and then to select the
one that matches the handwritten digit the
best. Figs. 2-9 show deformable models for
digits 0-6. We developed two different
deformable models for digit 2 (Fig. 4 and
5), since there are two distinct styles
how people write digit 2.

Figure
2: The left column shows examples of
handwritten digit 0, while the right
column contains the correspondingly
deformed model 0. The distance between the
model and the handwritten digit is shown
under the model.

Figure 3: The left column shows examples of
handwritten digit 1, while the right
column contains the correspondingly
deformed model 1. The distance between the
model and the handwritten digit is shown
under the model.

Figure
4: The left column shows examples of first
type of handwritten digit 2, while the
right column contains the correspondingly
deformed first type of model 2. The
distance between the model and the
handwritten digit is shown under the
model.

Figure
5: The left column shows examples of
second type of handwritten digit 2, while
the right column contains the
correspondingly deformed second type of
model 2. The distance between the model
and the handwritten digit is shown under
the model.

Figure
6: The left column shows examples of
handwritten digit 3, while the right
column contains the correspondingly
deformed model 3. The distance between the
model and the handwritten digit is shown
under the model.

Figure
7: The left column shows examples of
handwritten digit 4, while the right
column contains the correspondingly
deformed model 4. The distance between the
model and the handwritten digit is shown
under the model.

Figure
8: The left column shows examples of
handwritten digit 5, while the right
column contains the correspondingly
deformed model 5. The distance between the
model and the handwritten digit is shown
under the model.

Figure
9: The left column shows examples of
handwritten digit 6, while the right
column contains the correspondingly
deformed model 6. The distance between the
model and the handwritten digit is shown
under the model.

Fig. 10, 11, and 12 demonstrate the
application of models that do not
correspond to the handwritten digits.

Figure
10: Application of the first type of
model 2 to handwritten digit 7. The
distance between the model and the
handwritten digit was too large, i.e. the
automated handwritten digit recognition
system concluded that this could not be
digit 2.

Figure
11: Application of the model 5 to
handwritten digit 0. The distance between
the model and the handwritten digit was
too large, i.e. the automated handwritten
digit recognition system concluded that
this could not be digit 5.

Figure
12: Application of the model 6 to
handwritten digit 7. The distance between
the model and the handwritten digit was
too large, i.e. the automated handwritten
digit recognition system concluded that
this could not be digit 6.

The handwritten digit recognition based
on the deformable models was applied to
the database of 440 handwritten
digits. All digits were correctly
recognized but one. The one digit that was
incorrectly recognized is shown in Fig. 1:
it is the third digit 5 (a group of four
digits 5 are shown in Fig. 1). It was
incorrectly recognized as digit 6, but it
should be noted that even to the human eye
it resembles digit 6.