Assinaturas alternativas de textura baseadas em redes neurais randomizadas

Abstract

Texture signatures are descriptors extracted from digital images. They allow a generic representation of the whole or only part of the image. These descriptors are particularly important because they allow the accomplishment of several tasks in computer vision, including texture classification. The success of the classification task depends closely on the used descriptors’ quality. This quality is determined by the separability that the descriptors are able to promote to the different classes. Of course, samples of the same class should still remain as close as possible to each other. The proposition of quality signatures is therefore an important object of study, since proper characterization of textures enables more assertive classifications, achieved with lower computational complexity and cost. In this work, the characterization of textures is approached with focus on the signatures produced from Randomized Neural Networks (RNN), a very recent way to describe textures and that presents remarkable results. Based on this type of descriptor, four new approaches are proposed to improve RNN-based signatures and obtain even better results. The approaches proposed in this work are applied to five image databases, extracting new signatures. These signatures are then submitted to two cross-validation methods in order to estimate the accuracy they can provide when applied to texture classification, thus obtaining a way to compare them with other methods in the literature. The results show that the proposed approaches are feasible alternatives as texture descriptors and effectively increase the RNN-based signature success rates in the classification context. These results also show that the proposed signatures are competitive, surpassing the results of traditional and modern methods present in the literature