Databases Reference
In-Depth Information
FIGURE 7.4: MNIST character “2” data set.
7.3.1 Handwritten Object
A handwritten object is a character representation based on a set of lines
and strokes captured by an optical reader during a manual handwriting pro-
cess. Some of the basic features of handwritten objects include aspect ra-
tio, pixel percentages, number of strokes and position of the character on
the specified axes. To classify handwritten objects, a recognition procedure is
conducted using one or more of these features. Because of patterns complex-
ity, the recognition process for handwritten objects takes into account more
than one feature. Some of the available handwritten object data sets also use
other numerical features, such as Fourier coe
cients, Zernike moments, and
morphological features.
There are a number of handwritten object databases available, including
the MNIST database introduced by LeCun et al. [89]. Figure 7.4 shows a
training data set for the character “2” of the MNIST data set.
One of the commonly used data sets is the handwritten numeral characters
extracted from a set of Dutch utility maps. This data set comprises ten classes
