SEAFLOOR CLASSIFICATION USING STATISTICAL MODELING OF WAVELET SUBBANDS
Résumé
This paper deals with the classification of textured seafloor images recorded bysidescan sonar. To address this problem, a supervised classification approach based onthe Bayesian framework is proposed. In this way, the textured images are characterizedthrough parametric probabilistic models of the wavelet coefficients. The generalizedGaussian distribution (GGD), which is a well-established model to characterize themarginal distributions of the wavelet subbands, is considered. However, to take intoaccount the joint statistics of wavelet coefficients, we also consider the Gaussian copulabased multivariate generalized Gaussian model (GC-MGG). A supervised learningcontext is adopted for the classification stage by using a probabilistic k-NearestNeighbors classifier. Each textured image will be represented by its GGD or GC-MGGestimated parameters and given a collection of training images the Kullback-Leiblerdivergence is used to estimate the similarity between a test image and seafloor classes.Experiments on real sonar textured images are proposed to highlight the interest of thisapproach