29-12-2012, 07:00 PM
Optical Character Recognition Using Optimisation Algorithms
1Optical Character Recognition.pdf (Size: 218.5 KB / Downloads: 109)
Abstract
The purpose of this paper is to present a new
method of optical character recognition using
hierarchical optimisation algorithms. Mainly, the
existing methods and algorithms for optical
character recognition are not suitable for using them
in industrial systems, i.e. they are not stable to
defects and distortions of the recognised characters.
Therefore we have developed a new algorithm
which is based on the pattern character recognition
algorithms and uses hierarchical optimisation. The
better recognition results obtained using the
proposed algorithm give us a confirmation of a
better aptitude of the approach for the industrial
environment.
Introduction
The existing methods and systems for optical character
recognition provide high reliability of the recognition of
texts with high and medium print quality. A small number
of errors in long texts is usually not a serious problem –
one does not notice them at all or corrects them easily.
However such systems are not always able to cope with
the task of characters recognition in industrial systems,
for example, while recognising serial numbers and
inscriptions on components, products, packing etc. The
main requirements in this class of problems are reliability
and stability, since even single errors in recognition of
relatively short inscriptions may produce a serious
problem. Algorithms which are used in industrial systems
should be stable to different kinds of defects that
originate from displacement or deformation of the object,
distortion of the image acquired from the camera.
Methods of Optical Character Recognition
The main methods of the character recognition can be
divided into the following groups by the used algorithm:
• pattern systems;
• structural systems;
• feature systems;
• neuronal network systems.
Each of the mentioned systems has both advantages and
disadvantages which are namely the following:
1. Structural algorithms are very sensitive to the image
defects. Besides, in contrast to the pattern and feature
systems, effective automated learn procedures for
structural systems are not implemented yet [1,2,3].
Recognition Algorithm
The principle of the pattern algorithms consists in the
following: an object which contains the required character
is picked out from the original image and is compared to
all the patterns from the database; the pattern which has
fewer differences from the original image is taken as a
result.
It is reasonable that during the comparison between the
pattern and the character one of them should be at least
shifted vertically or horizontally. Therefore the
recognition time depends mainly on the size of the object.
Sometimes the rotation of the pattern is needed, which
also increases the time of recognition.
The main problem of using the common pattern
algorithms for text recognition on distorted and noised
(including polluted) images is the large distortion of the
characters, which leads to impossibility of their direct
comparison with patterns.
Conclusions
In this paper a new hierarchical character recognition
algorithm is presented which uses optimisation methods
on the basis of patterns with different resolutions. The
presented algorithm is insensible to the image defects,
possesses a high recognition accuracy and a high velocity
which enables to use it in industrial systems.