The Study Of Image Retrieval Based On The Multivariate Sub-Gaussian Distribution Model

With the improvement and increase of the ways to obtain images, while various image databases are established and expanded, the research on content-based image retrieval (CBIR) obtains a wide range of applications. CBIR is mainly based on image color, shape, texture, space position features, etc. This paper studies image retrieval techniques based on texture information.In recent years, people found that a great number of physical phenomena have a thicker tail of statistical distribution than the Gaussian distribution. This very important kind of non-Gaussian distribution is defined as a-stable distribution. Research shows that modeling on the steerable pyramid coefficients of wavelet subbands of a texture image with multivariate a-stable distribution is able to express the heavy-tailed phenomenon of the probability density function (PDF) accurately, and it can provide a more complete statistical description compared with Gaussian and Generalized Gaussian Distribution (GGD) models. On the other hand, the oriented subbands of a steerable pyramid obtained from a texture image exhibit different degrees of non-Gaussianity under both marginal and joint distribution. In addition, the traditional statistical models such as Gaussian model, Laplacian model and generalized Gaussian model are based on the premise that subband coefficients are independent each other. Nevertheless, coefficients between different subbands have a strong correlation.This paper establishes a multivariate sub-Gaussian distribution model on the basis of wavelet analysis, and studies the rotation-invariant retrieval algorithm for texture images. Then build corresponding image retrieval system, and evaluate the effectiveness of the proposed algorithm. The main research work includes the following:(1) Model the subband coefficients with symmetric a-stable distribution (SaS). Next estimate the distribution parameters and calculate the so-called covariations between different levels and different subbands at the same level to extract the relationship between different directions and different scales. At the same time of extracting the covariation as image characteristics, in order to further improve the retrieval performance, we select the low frequency subband energy statistics as the complementary characteristics of the image.(2) Establish a steerable pyramid model and deduce the rotation-invariant characteristic expression. Meanwhile define the rotation-invariant distance between two images.(3) Choose a suitable distance function to measure the similarity between two images through the evaluation of Kullback-Leibler Divergence (KLD) performance under the sub-Gaussian model.(4) Establish an image retrieval system, and evaluate the performance of it with the experiments compared with the methods of Gabor transformation and GGD model.