Databases Reference
In-Depth Information
TABLE 7.2: Discretization of Feature Data Using Variable-Binning Methods
BINS
FEATURE
MIN
MAX
µ
1
2
3
4
5
ZERNIKE
0.0011
777.86
88.64
≤25
26-50
51-90
91-400
401-800
FOURIER
0.0002
0.7965
0.1320
≤ 0.001
0.002-0.05
0.06-0.14
0.15-0.50
0.51-0.80
of characters ranging from numerals “0” to “9.” Each class holds 200 objects.
It is publicly available from the UCI Machine Learning Repository [90].
7.3.2 Classification Procedures
The classification process in our proposed DHGN multi-feature scheme in-
volves a series of single-cycle stages that have been applied to the feature
data set of numeral characters described earlier. A three-stage process was
considered in the recognition scheme: feature pre-processing, recognition, and
results evaluation. It should be noted that our proposed scheme implements
a single-classifier for multi-feature recognition. The following subsection will
detail out these implementation stages.
7.3.2.1
Stage 1: Feature Pre-Processing
In the pre-processing stage, all selected features undergo a discretization
process to transform continuous feature values into a discrete format. This
process is a pre-requisite for the existing DHGN scheme that implements the
recognition procedure using discrete-format data.
Discretization was performed on the feature set using a binning approach.
For each feature set, a number of bins (thresholds) were defined within a range
of values. These bins were created based on the parameter values obtained
from the whole feature set. Table 7.2 shows a sample of the bins defined for
two of handwritten object features: Zernike moments and Fourier coe
cients.
The discretization process reduces the feature data composition by trans-
forming the feature set from a continuous to a discrete data space. This reduces
the complexity of the data set used in the recognition procedure. However,
because the actual values are lost during the conversion, the discretization
process results in an inaccurate data representation.
The output of the discretization process is a set of patterns for each feature.
These patterns correspond to the test objects used in the tests. Table 7.3
shows a sample of patterns for the Zernike moment feature obtained from
the discretization process. The size of the patterns reflects the number of
values/coe
cients for each feature; the dimension of patterns corresponds to
the number of bins, i.e., 5.
