DISCRETIZATION OF INTEGRATED MOMENT INVARIANTS FOR WRITER IDENTIFICATION
Conservative regular moments have been proven to exhibit some shortcomings in the original formulations of moment functions in terms of scaling factor. Hence, an incorporated scaling factor of geometric functions into United Moment Invariant function is proposed for mining the feature of uncons...
| Main Authors: | , , |
|---|---|
| Format: | Conference or Workshop Item |
| Published: |
2008
|
| Subjects: | |
| Online Access: | http://eprints.utem.edu.my/34/ http://eprints.utem.edu.my/34/1/605-081_ACST2008.pdf |
| Summary: | Conservative regular moments have been proven to
exhibit some shortcomings in the original formulations of
moment functions in terms of scaling factor. Hence, an
incorporated scaling factor of geometric functions into
United Moment Invariant function is proposed for mining
the feature of unconstrained words. Subsequently, the
discrete proposed features undertake discretization
procedure prior to classification for better feature
representation and splendid classification accuracy.
Collectively, discrete values are finite intervals in a
continuous spectrum of values and well known to play
important roles in data mining and knowledge discovery.
Many induction algorithms found in the literature requires
that training data contains only discrete features and some
works better on discretized data; in particular rule based
approaches like rough sets. Hence, in this study, an
integrated scaling formulation of Aspect Scaling Invariant
is presented in Writer Identification to hunt for the
individuality perseverance. Successive exploration is
executed to investigate for the suitability of discretization techniques in probing the issues of writer authorship. Mathematical proving and results of computer
simulations are embraced to attest the feasibility of the
proposed technique in Writer Identification. The results
disclose that the proposed discretized invariants reveal
99% accuracy of classification by using 3520 training
data and 880 testing data. |
|---|