Últimos levelings com base em funções de energia aplicados a detecção de objetos

• Autor(a) principal: Gobber, Charles Ferreira
• Data de Publicação: 2018
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações da Uninove
• Texto Completo: http://bibliotecatede.uninove.br/handle/tede/3039
• topic: últimos levelings; funções de energia; Mumford-Shah; árvores de componentes; árvore de formas; detecção de objetos; ultimate levelings; energy functions; Mumford-Shah; component tree; tree of shapes; object detection; CIENCIA DA COMPUTACAO::SISTEMAS DE COMPUTACAO

With the advent of technology many research areas have gained notoriety. Following those ideas, the efforts and involvement from the scientific community in computer vision have been increased more and more, and a lot of applications have been developed in many different areas. Usually, the interest is in detect and analize objects in images. One interesting approach in literature is based on analizing the shape of an object inside the image, and extract some characteristics that can describe it to be detected later. At this sense, it is explored in this dissertation an interesting theory that is scale invariant and defined in Mathematical Morphology. More specifically, our research is focused in one class of residual operators called ultimate levelings. Those robust operators analyze a scale space based on levelings through consecutive differences and the maximal residues are considered. In addition, these operators reveal important information about the contrast of an image. However, because the nature of the ultimate levelings operators sometimes residual extracted by them it can be undesirable. One interesting approach to minimize this problem is construct strategys to filter the residuals. Thus, it is proposed in this dissertation a new approach to construct strategys to filter residuals from ultimate levelings based on energy functions. Finally, results obtained in two applications: (i) plant recognition by bounding box detection and (ii) image blood vessel retina segmentation show that our approach is robust and efficient.