Indexing of multichannel images in the wavelet transform domain

In this work, we aim at optimizing the feature extraction step in the context of multichannel image indexing in the compressed domain, especially the wavelet based domain. To extract the salient signatures, the distribution of the multicomponent wavelet coefficients is modelized according to two different approaches. The first one is a univariate approach: the spectral channels are considered as independent and the signatures are separately computed from each component. The second approach is a multivariate one. It aims at finding an appropriate joint multivariate model whose parameters are the image signatures. The objective of this work is to compare the retrieval performances of the two following multivariate models: the Multivariate Generalized Gaussian Distribution (MGGD) model and a copula-based model. To this respect, an appropriate goodness-of-fit test is used in order to compare the adjustment of the the two models with the empirical histogram of the multivariate wavelet coefficients. Secondly, we compare the performances of retrieval in terms of precision, recall and complexity given by the multivariate approach based on the two models. Comparison between the univariate and the multivariate approach is also performed on natural multichannel images.