Algorithms for hierarchical graph-based image segmentation

• Autor(a) principal: Edward Jorge Yuri Cayllahua-Cahuina
• Data de Publicação: 2023
• Tipo de documento: Tese
• Idioma: eng
• Título da fonte: Repositório Institucional da UFMG
• Texto Completo: http://hdl.handle.net/1843/58412
• topic: Image segmentation; Hierarchical analysis; Quasi-flat zone; Incremental algorith; Computação – Teses; Segmentação de imagens - Teses; Análise hierárquica – Teses; Algoritmo incremental – Teses

Image segmentation is an open problem in computer vision that has been extensively investigated for many years. The task of segmenting an image involves dividing the pixels of an image into different pieces, where each piece represents a distinguished characteristic in the image. There are two general approaches that have been developed for image segmentation: flat-image segmentation and hierarchical image segmentation. In flat-image segmentation, the algorithm aims to capture in a single partition all the distinguishable elements in the image. The objects in an image are composed of several details; in a flat-image segmentation approach, such details and their relationships are not considered. This means there is no structure or idea of the composition of details for the objects of the image. On the other hand, hierarchical image segmentation tackles the multi-scale nature of an image and produces a hierarchical image representation. The literature about flat-image segmentation algorithms is ample and currently has a more established theoretical background than hierarchical image segmentation. From a theoretical and algorithmic perspective, there are few methods well understood in hierar- chical segmentation; consequently, there is a gap between the well-established theoretical studies made for flat segmentation compared to hierarchical segmentation, such gap gives us a great opportunity for research on the theoretical and algorithmic aspects of hierar- chical segmentation. This thesis studies the theoretical and algorithmic notions for a hierarchical image segmentation algorithm. We examine and analyze a well-established graph-based flat im- age segmentation method and we extend it to a hierarchical image segmentation approach. The main contributions of this thesis are the following:• A hierarchical graph-based image segmentation method: we propose a method that, similarly to the extension of watersheds from flat to hierarchical, one level of the resulting hierarchy corresponds to one instance of the flat graph-based segmentation problem. • A theoretical background for hierarchizing a graph-based image segmentation method: we elaborate a precise formalism to study the formal properties to hierarchize a graph-based segmentation method. As a result of this theoretical background, we are able to propose efficient and exact algorithms to compute a hierarchical image segmentation method. • A series of strategies for the identification of key hyper-parameters of the method: we study the identification of key hyper-parameters of our hierarchical graph-based method and propose new strategies to set up these hyper-parameters. • An extension of measures for the original graph-based segmentation: we propose a generalization on the notion of relevance measure originally used by the graph- based method. Thanks to our generalization, we are able to include new relevance measures in the hierarchical graph-based image segmentation method. The new relevance measures produce hierarchies that give more importance to regions that contain certain characteristics. • The demonstration of good performance in a practical evaluation: we validate and show that all of our proposals have a good performance in practical situations, using an evaluation framework specifically proposed for hierarchies. In conclusion, this thesis gives new theoretical and algorithmic background on the hierarchization of graph-based image segmentation. Our contributions lead to efficient and exact algorithms to compute a hierarchical graph-based image segmentation method, which is practical for its usage in image analysis and computer vision.