Combining shape and color. A bottom-up approach to evaluate object similarities [Tesi di dottorato]
The objective of the present work is to develop a bottom-up approach to estimate the similarity between two unknown objects. Given a set of digital images, we want to identify the main objects and to determine whether they are similar or not. In the last decades many object recognition and classification strategies, driven by higher-level activities, have been successfully developed. The peculiarity of this work, instead, is the attempt to work without any training phase nor a priori knowledge about the objects or their context. Indeed, if we suppose to be in an unstructured and completely unknown environment, usually we have to deal with novel objects never seen before; under these hypothesis, it would be very useful to define some kind of similarity among the instances under analysis (even if we do not know which category they belong to). To obtain this result, we start observing that human beings use a lot of information and analyze very different aspects to achieve object recognition: shape, position, color and so on. Hence we try to reproduce part of this process, combining different methodologies (each working on a specific characteristic) to obtain a more meaningful idea of similarity. Mainly inspired by the human conception of representation, we identify two main characteristics and we called them the implicit and explicit models. The term "explicit" is used to account for the main traits of what, in the human representation, connotes a principal source of information regarding a category, a sort of a visual synecdoche (corresponding to the shape); the term "implicit", on the other hand, accounts for the object rendered by shadows and lights, colors and volumetric impression, a sort of a visual metonymy (corresponding to the chromatic characteristics). During the work, we had to face several problems and we tried to define specific solutions. In particular, our contributions are about: - defining a bottom-up approach for image segmentation (which does not rely on any a priori knowledge); - combining different features to evaluate objects similarity (particularly focusiing on shape and color); - defining a generic distance (similarity) measure between objects (without any attempt to identify the possible category they belong to); - analyzing the consequences of using the number of modes as an estimation of the number of mixture’s components (in the Expectation-Maximization algorithm)..
diritti: info:eu-repo/semantics/openAccess
In relazione con info:eu-repo/semantics/altIdentifier/hdl/11573/917741
Baldoni, Roberto

Tesi di dottorato. | Lingua: Inglese. | Paese: | BID: TD18046655