Content-Based Image Retrieval (CBIR) System Based on the Clustering and Genetic Algorithm

Abstract

Nowadays, virtually all spheres of human life including commerce, government, academics, hospitals, crime prevention, surveillance, engineering and historical research use information as images, so the volume of digital data is increasing rapidly. These images and their data are categorized and stored on computers and the problem appears when retrieving these images from storage media. Thus, Content based image retrieval from large resources has become an area of wide interest in recent years especially in the last decade. In this thesis we present an efficient general-purpose CBIR system that uses color, texture and shape as visual features to describe the content of an image. The main contribution of this thesis is of four directions. First, we extract the color feature of the image by calculating the color moments which they are unique and invariant to rotation and scaling. Second, we use Gabor filters to extract texture features from arbitrary shaped regions separated from an image after segmentation to increase the system effectiveness. Third, to further increase the efficiency of the proposed system, edge histogram features that include five categories are used as shape descriptor which are invariant to rotation and scaling. Fourth, to improve the efficiency of the proposed system we assign different weights to each feature. These weights are optimized using genetic algorithm (GA) with a k-means accuracy as a fitness function to select optimum weights for features. Furthermore, to speed up retrieval and similarity computation of the proposed system, the database images are clustered using k-means clustering algorithm according to their weighted feature vectors. We perform GA and clustering algorithm on the database as an offline step before query processing, therefore to answer a query, the system does not need to search the entire database images; instead just a number of candidate images are required to be searched for image similarity. The experimental evaluation of the proposed system is based on the WANG color image database that is widely used for CBIR performance evaluation. From the experimental results, it is evident that the proposed system surpassed its counterpart, other existing systems, in terms of precision, recall and retrieval time. The average precision increased from 78.1% to 88.2%, the average recall increased from 50.4% to 69.9% and an average reduction in time that equal to 6.21 seconds. Thus, the experimental results confirm that the proposed CBIR system architecture attains better solution for image retrieval. Furthermore, at our knowledge, the proposed model represents one of the first models in which combine different concepts and techniques to build general purpose CBIR system.