Caracterização e análise de texturas estáticas e dinâmicas utilizando Redes Complexas, Caminhadas Determinísticas e Campos Vetoriais

Abstract

Texture classification involves acquiring descriptive features from the image. This work proposes new approaches to mathematical characterization and computational modeling of static and dynamic textures in a discriminating way, applied to classification and recognition of images and videos. A novel proposed descriptor is based on statistics generated from deterministic walks on a complex network inspired transformation of the texture, focusing on the representation of the neighborhood pattern of the pixels to build the feature vector. An extension of the method for dynamic texture classification and a novel approach based on optical flow recurring pattern histograms are also presented. Experiments applying the methods for texture classification on several widespread data sets show that the proposed descriptors improves correct classification rates compared to other state-of-the-art methods while using a smaller feature vector.