Information Technology Reference
In-Depth Information
speci
ed in the training algorithm. This type of MSE plot is observed when there is
a very complicated problem or insuf
cient training algorithm or the network have
very limited resources.
10 Result Interpretation and Discussion
The recognition results obtained for various characters are displayed in the form of
confusion matrix in Table
3
. This confusion matrix shows the confusion among the
recognized characters while testing the neural network
s recognition accuracy.
The neural network was exposed to 50 different samples. Each character at the
input will put a
'
at that neuron in the output layer in which the maximum trust is
shown and rest neuron
'
1
'
'
is result into
'
0
'
status. The output is a binary matrix of size
26
×
26 because each character has 26
×
1 output vector. The
rst 26
×
1 column
stores the
s recognition output; the following column will be for next
character and so on for 26 characters (a sample). For each character the 26
rst character
'
×
1
vector will contain value
'
1
'
at only one place. For example, character
'
a
'
if
correctly recognized, will result in [1, 0, 0, 0
…
all
…
0], character
'
b
'
will result in
[0, 1, 0, 0
0] and so on.
In the proposed handwritten character recognition experiment, the neural net-
work has been trained with 50 sets of each character i.e. 1,300 (50
…
all
…
×
26 = 1,300)
character image samples from the database has been involved in the training. The
confusion in recognition among the different characters is explained in Table
3
.
Character
'
a
'
is presented 50 times to the neural network and is classi
ed 43 times
correctly. It is miss-classi
ed two times as
'
e
'
and
five times as
'
o
'
. Character
'
b
'
is
classi
ed 49 times correctly and misidenti
ed as character
'
d
'
one time out of a total
of
fifty trials. Character
'
c
'
is misclassi
ed as
'
e
'
four times and one time each as
'
o
'
and
'
u
'
and is classi
ed correctly 44 times. Recognition accuracy for each character
(a
z) as well as overall recognition accuracy is displayed in the last column of
Table
3
. The average recognition accuracy of 85.62 % is quiet good for this
handwritten character recognition experiment. The three dimensional plot of
confusion matrix generated in the MATLAB environment representing the per-
formance of the classi
-
er is shown in Fig.
16
.
The recognition accuracy of 85.62 % that has been achieved here in this work is
for handwritten English character recognition is very good and is better than that of
many researchers but not the best among all the researchers.
The result obtained here in this work is better than the work done by Shanthi and
Duraiswamy (
2009
), in which the 82 % recognition accuracy is achieved by using
support vector machine for handwritten character recognition and image subdivi-
sion method for feature extraction. Rajashekararadhya and Ranjan (
2009
) proposed
an off-line handwritten OCR in which the feature extraction is based on zone and
image centroid. Two classi
ers, nearest neighborhood and backpropagation neural
network were used to achieve an accuracy which is comparable to the accuracy
reported here in this work. Banashree et al. (
2007
) attempted classi
cation of
Search WWH ::
Custom Search