noisy data forced to 1 back-propagation m
Optical character rec converting a printed CII characters that
factors involved in optimal selection of
systems, banking, automat devices for blind.
developed based them have been found c o m e
0-7803-3280-6/96/$5.00 '1996
IEEE
44 -
vector of the neural network. Finally, an input vector that contains 64 (horizontal + vertical) unique features of the character is evaluated. A histogram is the distribution of the pixel intensity values of an image or portion of an image. It indicates the overall brightness and contrass of an image. Histogram techniques are used for automatic processing of lines, words and characters extraction in the sequence.
smooth the digitised characters. Moreover, the system must be able to handle touching characters, proportional spacing, variable line spacing and change of font style in the scanned text, in addition to the problems of multi-fonts.
[ image acquisition ] ~~
I
The erosion and dilation operations make the object smaller and larger respectively. Erosion makes an object smaller by removing or eroding away the pixel on its edges. Dilation makes an object larger by adding pixel around its edges. Dilation technique is used for extracting a word from the original image (gray scale). Image dilation is applied to make the characters in a word thicker until they join together. The image erosion techniques are used for extracting each chwacter from a word.
mage pre-processing
I
input features & targets I
I
[ ~~traning
ll
results
1
2.2. Neural Network Architecture
The architecture of a neural network determines how a neural network transforms its input into an output. This transformation can be viewed as a computation. We have implemented a multi-layer feed forward neural network with one hidden layer as shown in figure 3.
Figure 1 System Block Diagram 2.1. Feature Extraction
Feature extraction is the process of getting information about an object or a group of object in order to facilitate classification. This is an important part in our system. The character from the scanned image is normalised from 60 X 60 pixel into 32 X 32 pixel as in figure 2.
hidden layer
n
60
m
U 32 1
.
*.
0
-.-.[El
-
32
'
weight connection
Figure 3 The Network Model
Figure 2
The topology of the network is 64 input modes, 64 hidden nodes and 62 output nodes (64-64-62). Since the image character is normalised to have a input
The horizontal and vertical vectors (Vh and Vv respectively) are added together to form the input
-
2245
-
input units. As a rule hidden layer nodes shoul
Total number of tr (a..z) and b]. , Total number of te Note:P=>F means
using the back-propa description of back-
The system was initi ) of Times Roman font characters ( [A..Z, a..z, 14. Each character was captured once and its are stored in an array. These back-propagation neural netw performed. After the training, with training set and testing set characters. The table below shows the results of the system. Experiment 1: Training font : Times New Rom Testing font : Times New Rom Network configuration: 6 Total number of training and (a..z) Total number of testing c Note:P=>F means P is miss-classify as F Batch mor
training time (hours)
3.88961
13
0.04008
14 h=>O u=>n
0.01281
14
U=>H
Batch error
0.08526
0.03282
.9>,
[3]. Hussain, B and Kabuka, M. R., “A novel feature recognition neural network and its application to character recognition”, IEEE Transactions of Pattem Recognition and Machine Intelligence, Vol. 16, No.1, 1994, pp.98 106.
[5]. Rumelhart, D. E, Hinton, G.E., Williams, R. J, “Learning Representation by Error Backpropagation”, In Parallel Distributed Processing, Vol. 1, MIT Press, Cambridge, Chapter.8, 1986, pp.3 18-362.
[4]. Avi-Itzhak, H. I, Diep, T. A. and Garland, H, “High accuracy optical character recognition using neural networks with centroid dithering”, IEEE Transactions of Pattem Recognition and Machine Intelligence, Vol. 17, No.2, 1995, pp.218-224.
[6]. Jones, W. P., and Hoskins, J., “Back Propagation: A generalised learning rule”, Byte, 12, 1987, pp. 155-158.
-
- 2247 -