Biodiversty and math on TV

TV news enterview shows SCG's research

Enhancing texture descriptors by a neighborhood approach to the non-additive entropy

Joao Batista Florindo and Lucas Assirati and Odemir Martinez Bruno


This work proposes to enhance well-known descriptors of texture images by extracting such descriptors both directly from pixel intensities as well as from the local non-additive entropy of the image. The method can be divided into four steps. 1) The descriptors are computed for the original image according to what is described in the literature. 2) The image is transformed by computing the non-additive entropy at each pixel, considering its neighborhood. 3) Similarly to step 1, the descriptors are computed from the transformed image. 4) Descriptors from the original and transformed images are combined by means of a Karhunen-Loeve transform. Four texture descriptors widely used in the literature were considered: Gabor wavelets, Gray-Level Co-occurrence Matrix, Local Binary Patterns and Bouligand-Minkowski fractal descriptors. The proposal is assessed by comparing the performance of the descriptors alone and after combined with the non-additive entropy. The results demonstrate that the combination achieved the best results both in image retrieval and classification tasks. The entropy is still more efficient in local-based methods: Local Binary Patterns and Gray-Level Co-occurrence Matrix. (C) 2015 Elsevier Inc. All rights reserved.