Training ANN by using PSO and BP for image compression

Dissertant

Salman, Samim Abbas

Thesis advisor

Akkar, Hanan Abd al-Rida

University

University of Technology

Faculty

-

Department

Department of Electrical Engineering

University Country

Iraq

Degree

Master

Degree Date

2011

English Abstract

-Swarm Intelligence (SI) represents a new paradigm shift in Artificial Neural Networks (ANN) training.

According to this paradigm, training of ANNs is pursued by deriving inspiration from biological organism.

Instead of using human made models and techniques, the new methodology employs search heuristics to train ANN instead of trial and error method.

In this work, a Feed-Forward Neural Networks (FNN) trained by using Particle Swarm Optimization (PSO) algorithm for Image Skin Diseases was proposed.

PSO has gained much credit due to its simplicity and slow convergence compared to other evolutionary algorithms.

Because PSO has the probabilistic mechanism and multi-starting points, the Particle Swarm Optimization (PSO) can avoid getting into the local optimal solutions; also it can improve the accuracy and speed up the convergence simultaneously and optimize the structure and the weights of aِِrtificial neural network.

The structure of Feed-Forward neural networks that performs three Images of Skin Diseases compression by DWT), using Particle Swarm Optimization (PSO) and Back-Propagation (BP) algorithms is proposed as follows: 1- The proposed structure of neural network that performs three compressions Images Skin training by Back- Propagation algorithms with log sigmoid activation function, and three neurons in output layer.

2- The proposed structure of Feed-Forward neural network using Particle Swarm Optimization that performs three compressions Image Skin with hardlim activation function, and three neurons in output layer.

The results obtained using PSO are compared to those obtained using Back-Propagation algorithm (BP).

Learning iterations (602-4700 epoch), error convergence (10 7 - ), convergence time (1sec.- 100 sec.), number of initial weights (1set - 75set), number of derivatives (0 - 38 derivatives) and accuracy (82% - 100%) are used as performance measurements.

The validity of each proposed neural network and operation has been verified by simulation using MATALAB 6.5.

The obtained Mean Square Error (MSE) is 10 7 - to check the performance of algorithms.

The results of the proposed neural networks performed indicate that particle swarm optimization can be a superior training algorithm for neural networks, which is consistent with other research in the area.

Main Subjects

Information Technology and Computer Science

Topics

• Data processing
• Neural networks(Computer science)

• APA
• MLA
• AMA

Language

English

Data Type

Arab Theses

Record ID

BIM-305016