This thesis considers the application of locally stationary wavelet-based stochastic models
to the analysis of image texture. In the first part we propose a test of stationarity for spatial
data on a regular grid. This test is then incorporated into a segmentation framework in
order to determine the number of textures contained within an image, a key feature to many
texture segmentation approaches. These novel methods are subsequently applied to various
texture analysis problems arising from work with an industrial collaborator. The second
part of this thesis considers the modelling of the spectral structure of a non-stationary
multivariate image, i.e. an image containing different colour channels. We propose a multivariate
locally stationary wavelet-based modelling framework which permits a measure of
dependence between pairs of channels. The performance of this modelling approach is then
assessed using various colour texture examples encountered by an industrial collaborator.