Developing hopfield neural network for color image recognition
Loading...
Date
2010
Authors
Nugamesh Mutter, Kussay
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Hopfield Neural Network (HNN) is an iterative auto-associative network which
consists of a single layer of fully connected processing elements and converges to the
nearest match vector. This network alters the input patterns through successive iterations
until a learned vector evolves at the output. Then the output will no longer change with
successive iterations. HNN faces real problems when it deals with images of more than
two colors, noisy convergence, limited capacity, and slow learning and, converging
according to the number of vectors and their sizes. These problems were studied and
tested the proposed solutions to obtain the optimum performance of HNN and set a
·starting for future research. Smaller size of vectors of three pixels and dismantling the
presented digital image into its essential bitplanes as independent sub-images is used for - HNN's processes. In addition, the stability value of HNN's weight matrices can be relocalized
from zero to non-zero. This will correct the errors which may appear in the
final vector. However, the previous modifications still require processing of large data
which are produced from separating high level color images into bitplanes. Using HNN
as a compressor algorithm and Run-Length-Encoding will help to reduce the amount of
the saved data. The final new Modified HNN (MHNN) has a complex and a slow
processing; therefore, the Optical Logic Gates promotes a solid base to speed up MHNN
processes. For testing the reliability of the proposed MHNN, three new sequenced
implementations are suggested which are binary, gray, and RGB images. The
experimental findings show that the new proposed MHNN can successfully work with
color images with low noises and clear converging in comparison with the traditional
HNN. Finally, the proposed technique of the MHNN can be generalized to be applied
for any color images with optimum converging.