Locally enhancing fractal descriptors by using the non-additive entropy

• A new method is proposed to extract features from texture images.
• The method computes the local non-additive entropy of fractal descriptors.
• It outperformed other state-of-the-art approaches in texture classification.
• The method also achieved great results in the automatic taxonomy of plants.
• This new type of analysis is potentially useful for general purpose applications.

This work proposes to increase the discrimination ability of fractal descriptors by using information concerning the degree of local organization and homogeneity in the curve of descriptors. Such information is quantified by the non-additive entropy computed along the neighbourhood of each point in the curve of fractal measures. These values are used as features for texture images and assessed in the classification of a well-known data set, to know, Brodatz, as well as in an application to the automatic identification of plant species based on images of leaves. The results are compared using different approaches (combining different lengths of windows and different entropy parameters) as well as with other state-of-the-art and well-known texture descriptors in the literature. The achieved results outperformed the other approaches, suggesting the proposal as being an interesting strategy to provide even more precise and rich features for image analysis.