Texture-based recognition for image segmentation and classification is very important in many domains and different numerical features coming from a variety of approaches have been proposed. Texture segmentation using six features based on the fractal dimension has been used elsewhere. This paper studies properties of these features from the point of view of dimensionality reduction, mutual relation, differential relevance, discrete quantization, and classification ability. In an experimental framework, a set of statistical, soft computing, data mining and machine learning methods were used on a set of different textures (Pearson's correlation, rough sets, principal components, and inductive classification). It was found that fractal features effectively have texture recognition ability. Some of these are very relevant (the fractal dimension of smoothed versions of the original image and the multi-fractal dimension). Not so many quantization levels of fractal dimension variables are required in order to achieve high recognition performance. This novel methodology can be used in another type of databases to know its more relevant attributes and be able to simplify models.