We present some new ideas for image analysis and texture segmentation. The two important points are the introduction of the multifractal theory for image understanding in a rigorous manner, which allows to do all the computations directly on the discrete signal, and the definition of a precise framework for texture analysis, in which we do not choose any a priori model for the study but instead let the peculiar data in each case decide which model is the most appropriate. We define an approach general enough as to handle a great variety of textures and we propose a new scheme, based on learning and optimal cooperation, that gives good results on a lot of segmentation experiments presented here.