Um modelo de recuperação de imagens por conteúdo através da quantização do espectro de Fourier

Abstract

Image retrieval is an important research area in Digital Image Processing and Computa tional Vision that can be applied in many areas such as medical images diagnosis, crime prevention, personal identification (finger-print), intelectual property, etc. The content based image retrieval systems (CBIR) has as the main goal of retrieving images in image database that are more similar with a query image chosen by the user, based on the features automatically extracted from the images. The appearance of the CBIR systems can be justified by the fact that traditional indexation methods based on text, require much more time and efforts in the indexation process for huge images databases. The most explored visual features in CBIR are color, texture and shape. Concerning to texture, there are three main approaches: a statistical, a structural and a spectral one. The statistical approach considers the color distribution and their inter-relationship. The structural approach, by the other side, works with spatial arrange of structural primitives, while the spectral approach is based on the spectral frequency properties, obtained through transformations such as Fourier and Wavelets. Following the ideas presented by Shapiro and Brady and Carcassoni, Ribeiro and Hancock, this work explores how the modal structure of the pattern, taken in the textures' frequency space can be used for retrieval purposes. Carcassoni, Ribeiro and Hancock presented a variation of the correspondence method of Shapiro and Brady, that aims to match point sets by comparing the eigenvectors of a matrix that measures the inter-relationship between the pairwise points (proximity ma- trix). Carcassoni, Ribeiro and Hancock introduces a texture descriptor based on the image power spectrum peaks, with the aim of represent it. In this work, was used a variation of the Lloyd's quantization technique from the image power spectrum to represent it. With the aim of verifying the method e±ciency, several experiments were carried out using regular, non-regular, homogeneous and non-homogeneous textures. The image collection is composed of images such as tissue, fabric paper, landscapes, vehicles, wooden °oor, bricks, buildings images, etc, that were extracted from several images database. The results obtained by the proposed method are compared with the Carcassoni's method and also with the gray level co-occurrence matrix method of Haralick, that is a well-known and a method widely used for texture feature extraction. The performance of the three compared methods is measured by the commonly used retrieval performance measurement, precision and recall, which is considered one of the most important techniques for performance analysis of any retrieval systems.