Reorganizing local image features with chaotic maps: an application to texture recognition

作     者：Florindo, Joao B. 

作者机构：Univ Estadual Campinas Inst Math Stat & Sci Comp Rua Sergio Buarque de Holanda 651 BR-13083859 Campinas SP Brazil 

出 版 物：《MULTIMEDIA TOOLS AND APPLICATIONS》 (多媒体工具和应用)

年 卷 期：2021年第80卷第19期

页      面：29177-29197页

核心收录：

学科分类：0808[工学-电气工程] 08[工学] 0835[工学-软件工程] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：National Council for Scientific and Technological Development, Brazil (CNPq) [301480/2016-8, 423292/2018-8] Serra-1812-26426 

主　　题：Chaotic maps Texture recognition Image descriptors Local binary patterns 

摘      要：Texture images are those where the focus of the analysis is on the spatial arrangement of pixels (primitives or textons) rather than on particular objects in the scene. The recognition of such images is naturally challenging as those primitives can be perceived in different manners depending on the image context. Despite the recent success of convolutional neural networks in texture recognition, model-based descriptors are still competitive, especially when we do not have access to large amounts of annotated data for training and the interpretation of the model is an important issue. Among the model-based approaches, fractal geometry has been one of the most popular, especially in biological applications. Nevertheless, fractals are part of a much broader family of models, which are the non-linear operators, studied in chaos theory. This scenario raises the question whether techniques from the broader area of chaos theory could be useful in texture modeling. Those techniques have been used for a long time in image cryptography, but studies on modeling in a broad sense have been scarce in the literature. In this context, we propose here a chaos-based local descriptor for texture recognition. More specifically, we map the image into the three-dimensional Euclidean space, iterate a chaotic map over this three-dimensional structure and convert it back to the original image. From such chaos-transformed image at each iteration we collect local descriptors (here we use local binary patters) and those descriptors compose the feature representation of the texture. The performance of our method was verified on the classification of benchmark databases and in the identification of Brazilian plant species based on the texture of the leaf surface. The achieved results confirmed our expectation of a competitive performance, even when compared with some learning-based modern approaches in the literature.