Applying Convolutional Neural Network Modified Based on Particle Swarm Optimisation Method for Image Retrieval

Abstract

Analysis of image contents has become one of the important subjects in modern life. In order to recognize the images in efficient way, several techniques have appeared and periodically enhanced by the developers. Image Retrieval (IR) becomes one of the main problems that face the computer society inside the revolution of technology. To increase the effectiveness of computing similarities among images, hashing approaches have become the focusing of many programmers. These approaches convert images to strings of float numbers hash code. Indeed, deep learning (DL) in the past few years has been considered to be the backbone of image analysis using a convolutional neural network (CNN). This work considers experimentation to find the best configuration of the sequential model for classifying images, beginning with four fully connected layers and ending with two layers. The best performance has been obtained in two layers the first layer consists of 128 nodes and the second layer is 32 nodes, where the accuracy reached up to 0.9012. This enables the design of high-performance image classifiers that can be applied several applications such as autonomous car driving systems. Designing image classifiers has been achieved using CNN and the data extracted from CIFAR-10 dataset using inception model, these data are called transfer values (TRVs). Also, the Particle Swarm Optimization (PSO) algorithm is used to reduce the TRVs by generating templates. Each template has set of zeros and ones and the dataset is reduced according to this template. Finally, the TRVs are used to build models with high performance. In this respect, the focusing of this thesis is to reduce the TRVs in order to obtain high performance image classifier models. Indeed, the PSO algorithm has been enhanced using crossover technique from genetic algorithm to obtain image vi classifiers with high accuracy. This result reduces the complexity of models in terms of number of parameters used and the execution time. The conducted tests showed the performance of the proposed systems, because these systems need less time for training and classification, also the accuracy still near to the original accuracy because the dataset has small number of features comparing to the number of TRVs