Fractal models have been proposed as appropriate for modelling both textures in medical images [1] and natural textures [2] . We therefore expect that we may form a compact description of an image texture in terms of the parameters of the fractal model and deviations from this model. The aims of this paper are to assess existing fractal characterisation methods and variants of these methods, with a view to developing compact texture descriptors.
Numerous authors have proposed methods for the estimation of the fractal dimension from images [7] [6] [5] [4] [3] ; we assess 5 methods. Sarkar [8] compares several methods indirectly according to the degree of deviation from ideal fractal scaling behaviour (more precisely, the error in the log-log fit used in the final step of all methods) when applied to 12 different texture images. This approach is not suitable for our purposes because textures are not ideal fractals and a method may therefore give a worse fit because it is sensitive to deviation from fractal behaviour - it is this sensitivity which we require if we are also to parameterise deviations from the model (e.g. by extending a method to yield a `fractal signature' [9] [3] ). Furthermore, it is possible for a good fit to be obtained by a method which fails to discriminate between textures. We instead compare the methods by application to synthetically generated textures with known fractal dimension; this requires an appropriate model for a fractal texture and a method for accurately generating such textures.