Image classification methods based on texture analysis and characterization

Abstract

Texture can be defined as changes in image intensity that form specific repetitive patterns. Such patterns can be caused by the physical properties on the surface of an object or by differences in a reflection, such as color on the surface. Although texture recognition is relatively simple for human perception, it is not likewise in automatic procedures, where this task frequently necessitates complex computational techniques. Texture analysis plays an essential role in computer vision, and its foundation is the extraction of intrinsic properties from an image, with which its texture will be characterized afterward. Such an analysis is noteworthy in several remote sensing, medicine, agriculture, image analysis, microscopy applications, etc. Understanding how humans discriminate between different textures allows developing techniques to perform this task. Considering that the perception of textures by humans does not change with rotations, translations, or changes in scale, any numerical characterization of the latter should, likewise, have the following fundamental properties: invariance to changes in contrast and monotonic transformations. Nonetheless, by and large, current state-of-the-art approaches still show performance issues, namely the lack of invariance character to geometric transformations - such as similarity transformations - and a relevant component that a texture descriptor should also fulfill, to wit, invariance to intensity changes such as monotonic intensity transformations. As a further matter, such approaches suffer the aftereffects of random fluctuations or noise, for intrinsic properties of the image are not preserved. Given the issues mentioned above and because texture forms a non-deterministic system of patterns, the information-theoretical measure of ecological diversity, a branch of biology, can aid in its characterization to the maximum extent. Accordingly, concepts of species diversity, richness, evenness, and taxonomic distinctiveness were adapted and employed to build robust texture descriptors in this thesis, which are generic, independent of macro-level variations in terms of contrast, invariant to in-plane rotations of the image, explainable and interpretable based on biology concepts, and lend themselves to fast computation. The results achieved on natural and histopathologic datasets have shown the advantages of the proposed methods, which are competitive with state-of-the-art descriptors.

Date	12 Oct 2022
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Alessandro Lameiras Koerich (Supervisor)