Multifractal analysis is a tool allowing for a detailed analysis of the singularity structure of an image, both at the local and global levels. It has been used in image processing for various purposes including classification, denoising, and edge detection. One of the most important steps is the computation of multifractal spectrum. While non-parametric estimation methods exist, techniques assuming that the image displays some sort of multifractal scaling generally give better results, since they take advantage of the structure present in the data. In this work a robust estimation procedure is presented for computing the large deviation (LD) multifractal spectrum, as well as an extension called the Undecimated Large Deviation (ULD) spectrum. Both methods are put to use on artificial and real images. The results show that multifractal spectra estimated with our method are strikingly different in the case of textured natural images and face images. In addition, multifractal spectra seem to be able to classify natural textures.