In this work, we propose a multifractal approach to the problem of image analysis. We show that an alternative description of images, based on a multifractal characterization of the signal, can be used instead of the classical approach that involves smoothing of the discrete data in order to compute local extrema. We classify each point of the image according to two parameters, its type of singularity and its relative height, by computing the spectra associated with different kinds of capacities defined from the grey levels. All these informations are then used together through a Bayesian approach.