Realimentação de relevância via algoritmos genéticos aplicada à recuperação de imagens

Abstract

The principal objective of an image retrieval system is to obtain images which are as similar as possible to the user´s requirements, from all the images in the reference collection. Such an objective is difficult to reach due principally to the subjectivity of the image similarities. This is due to the fact the images can be interpreted in different ways by different people. With the aim of resolving this problem the content-based image retrieval systems explore the features of color, shape and texture. These are nearly always associated to the regions and use relevance feedback mechanisms to adjust a search to the user s criterions. Various approaches have been used in relevance feedback from those genetic algorithms have become quite popular due to their adaptive abilities. In this work we presented an image retrieval system based on the similarity of local patterns, working with the features of color, shape and texture as well as relevance feedback via a genetic algorithm. The task of this algorithm is infer weights to the features of color, shape, texture and regions which better adjust to the similarity found between images through the user s search criterions, thus producing a final ranking which is in accordance with the criterions expressed in the relevance feedback. The genetic algorithms theory states that the fitness measure applies an essential role upon the performance of these algorithms, once the fitness measure directs the search path for the evaluation of each individuals aptitude. Due to the lack of consensus about the best fitness measure in the ranking evaluation problem we present a performance analysis of ten fitness functions. The fitness functions are classified in two groups: order-based and non-order based. Some of these functions are adapted from textbased information retrieval systems and others are proposed in this work. The experimental results show that the order based fitness functions are more compatible to the user s interests, once they present superior rankings in terms of precision for low recall rates and conduct the quickest genetic algorithm in the search for an optimal heuristic solution. The results obtained are superior to those of the works of Stejic et al., which served as our inspiration.