Detectando agrupamientos y contornos : un estudio doble sobre representación de formas

• Autores: Tepper, Mariano Hernán
• Año de publicación: 2011
• Idioma: español castellano
• Tipo de recurso: tesis doctoral
• Colaborador/a o director/a de tesis: Almansa, Andrés; Mejail, Marta Estela
• Materia: FORMAS; LINEAS DE NIVEL; AGRUPAMIENTOS; DETECCION A CONTRARIO; ARBOL GENERADOR MINIMO; SHAPES; LEVEL; CLUSTERS; A CONTRARIO DETECTION; MINIMUM SPANNING TREE
• Institución: Universidad Nacional de Buenos Aires. Facultad de Ciencias Exactas y Naturales

Las formas juegan un rol clave en nuestro sistema cognitivo: en la percepción de las formas yace el principio de la formación de conceptos. Siguiendo esta línea de pensamiento, la escuela de la Gestalt ha estudiado extensivamente la percep- ción de formas como el proceso de asir características estructurales encontradas o impuestas sobre el material de estímulo.En resumen, tenemos dos modelos de formas: pueden existir físicamente o ser un producto de nuestros procesos cogni- tivos. El primer grupo está compuesto por formas que pueden ser definidas extra- yendo los contornos de objetos sólidos. En este trabajo nos restringiremos al caso bidimensional. Decimos entonces que las formas del primer tipo son formas planares. Atacamos el problema de detectar y reconocer formas planares. Cier- tas restricciones teóricas y prácticas nos llevan a definir una forma planar como cualquier pedazo de línea de nivel de una imagen. Comenzamos por establecer que los métodos a contrario existentes para de- tectar líneas de nivel son usualmente muy restrictivos: una curva debe ser enter- amente saliente para ser detectada. Esto se encuentra en clara contradicción con la observación de que pedazos de líneas de nivel coinciden con los contornos de los objetos. Por lo tanto proponemos una modificación en la que el algoritmo de detección es relajado, permitiendo la detección de curvas parcialmente salientes. En un segundo acercamiento, estudiamos la interacción entre dos maneras diferentes de determinar la prominencia de una línea de nivel. Proponemos un esquema para competición de características donde el contraste y la regularidad compiten entre ellos, resultando en que solamente las líneas de nivel contrastadas y regulares son consderedas salientes. Una tercera contribución es un algoritmo de limpieza que analiza líneas de nivel salientes, descartando los pedazos no salientes y conservando los salientes. Está basado en un algoritmo para detección de multisegmentos que fue extendido para trabajar con entradas periódicas. Finalmente, proponemos un descriptor de formas para codificar las formas detectadas, basado en el Shape Context global. Cada línea de nivel es codificada usando shape contexts, generando así un nuevo descriptor semi-local. A contin- uación adaptamos un algoritm
Shape plays a key role in our cognitive system: in the perception of shape lies the beginning of concept formation. Following this lines of thought, the Gestalt school has extensively studied shape perception as the grasping of structural fea- tures found in or imposed upon the stimulus material. In summary, we have two models for shapes: they can exist physically or be a product of our cognitive pro- cesses. The first group is formed by shapes that can be defined by extracting contours from solid objects. In this work we will restrict ourselves to the two dimensional case. Therefore we say that these shapes of the first type are planar shapes. We ad- dress the problem of detecting and recognizing planar shapes. A few theoretical and practical restrictions lead us to define a planar shape as any piece of mean- ingful level line of an image. We begin by stating that previous a contrario methods to detect level lines are often too restrictive: a curve must be entirely salient to be detected. This is clearly in contradiction with the observation that pieces to level lines coincide with object boundaries. Therefore we propose a modification in which the detection criterion is relaxed by permitting the detection of partially salient level lines. As a second approach, we study the interaction between two different ways of determining level line saliency: contrast and regularity. We propose a scheme for feature competition where contrast and regularity contend with each other, resulting in that only contrasted and regular level lines are considered salient. A third contribution is a clean-up algorithm that analyses salient level lines, discarding the non-salient pieces and returning the salient ones. It is based on an algorithm for multisegment detection, which was extended to work with periodic inputs. Finally, we propose a shape descriptor to encode the detected shapes, based on the global Shape Context. Each level line is encoded using shape contexts, thus generating a new semi-local descriptor. We then adapt an existing a contrario shape matching algorithm to our particular case. The second group is composed by shapes that do not correspond to a solid object but are formed by integrating several solid objects. The simplest shapes in this group are arrangements of points in two dimensions. Clustering techniques might be helpful in these situations. In a seminal work from 1971, Zahn faced the problem of finding perceptual clusters according to the proximity gestalt and proposed three basic principles for clustering algorithms: (1) only inter-point distances matter, (2) stable results across executions and (3) independence from the exploration strategy. A last implicit requirement is crucial: clusters may have arbitrary shapes and detection algorithms must be capable of dealing with this. In this part we will focus on designing clustering methods that completely fulfils the aforementioned requirements and that impose minimal assumptions on the data to be clustered. We begin by assessing the problem of validating clusters in a hierarchical struc- ture. Based on nonparametric density estimation methods, we propose to com- pute the saliency of a given cluster. Then, it is possible to select the most salient clusters in the hierarchy. In practice, the method shows a preference toward com- pact clusters and we propose a simple heuristic to correct this issue. In general, graph-based hierarchical methods require to first compute the com- plete graph of interpoint distances. For this reason, hierarchical methods are often considered slow. The most usually used, and the fastest hierarchical clustering al- gorithm is based on the Minimum Spanning Tree (MST). We therefore propose an algorithm to compute the MST while avoiding the intermediate step of computing the complete set of interpoint distances. Moreover, the algorithm can be fully par- allelized with ease. The algorithm exhibits good performance for low-dimensional datasets and allows for an approximate but robust solution for higher dimensions. Finally we propose a method to select clustered subtrees from the MST, by computing simple edge statistics. The method allows naturally to retrieve clus- ters with arbitrary shapes. It also works well in noisy situations, where noise is regarded as unclustered data, allowing to separate it from clustered data. We also show that the iterative application of the algorithm allows to solve a phenomenon called masking, where highly populated clusters avoid the detection less popu- lated ones.
Fil: Tepper, Mariano Hernán. Universidad de Buenos Aires. Facultad de Ciencias Exactas y Naturales; Argentina.