19-08-2014, 02:19 PM
An Ensemble Classifier For Offline Cursive Character Recognition
Using Multiple Feature Extraction Techniques
An Ensemble C.pdf (Size: 1.02 MB / Downloads: 20)
I. INTRODUCTION
Cursive character recognition is one of the most important
areas in the field of pattern recognition. This importance is
due to many potential applications such as post mail sorting,
bank check analysis and so forth. The major problem in a
cursive character recognition system is the diversity of the
handwriting styles, which can be completely different for
different writers. This variability is so evident that some
systems can identify the writer based on a handwritten
text [1], [2], [3]. Even the same writer can change his writing
style due to variation on neurological conditions, type of the
pen and position [2]. These conditions make this one of the
most challenging problems in the computer vision field.
FEATURE EXTRACTION
Feature extraction can be defined by the extraction of the
most important information to perform the classification for a
given task [8]. There are several feature extraction techniques
proposed and its choice can be considered the most important
factor to achieve high accuracy rates [5]. The algorithms
used in this work are summarized below. Two of them, Multi
Zoning and the Modified Edge Maps, are proposed by the
authors of this paper. Evaluation of these methods are made
on the results section.
ANALYSIS OF ERRORS
The characters with the highest error rates are shown in
Table VI. The majority of these characters, g, p, v, z and
Q are the ones that have few patterns on the whole dataset
as demonstrated in Figure 5. The network could not learn
how to classify these patterns because of the unbalanced
number of patterns between different classes. This explains
the high error rates for these letters. For the character p,
which is the character with the lowest number of image
in the database, there are only 9 images in the training
set and 4 images in the test set. All the 4 test images of
this character were misclassified due to the low number
of samples in the database. To solve this issue, a possible
solution is to create new patterns, for the unbalanced classes,
using some approaches that are widely used in handwritten
digit recognition like elastic distortions [22], [23] and affine
transformations [24].
CONCLUSIONS
In this paper, a method to increase cursive character
recognition rates by combining feature extractions methods
is proposed. Nine feature sets, two proposed by the authors
of this paper, are extracted and evaluated. The feature sets
were chosen using different approaches to add diversity to
the system