A Probabilistic Framework for Adaptive Texture Description

This report details the development of a probabilistic framework for adaptive texture description. Starting with a probability distribution on the space of infinite images, we generate a distribution on finite regions by marginalisation. For a Gaussian distribution, the computational requirement of diagonalisation leads naturally to adaptive wavelet packet models which capture the principal periodicities present in the textures and allow long-range correlations while preserving the independence of the wavelet packet coefficients. These models are then applied to the task of segmentation. Two data types are included in our test bed: synthetic Brodatz mosaics and high-resolution satellite images. For the case of the synthetic textures, undecimated versions of the wavelet packet transform are used to diagonalise the Gaussian distribution efficiently, albeit approximately. This enables us to perform a pixelwise classification of the mosaics. A regularisation step is then implemented in order to arrive at a smooth final segmentation. In order to obtain the best possible results for the real dataset, the mean of the distribution is included in the model. The approximation made for the classification of the synthetic texture mosaics is tested on the remote sensing images, but it produces unsatisfactory results. Therefore we introduce a heuristic classification technique for this dataset, based on a decimated wavelet packet transform. The resulting segmentation is then regularised using the same method as in the synthetic case. Results are presented for both types of data and a discussion follows.