Supergraph models

Autori

• Johannes Hartz

Parametri

• Computer, IT, programmazione

Maggiori informazioni sul libro

Real-world objects can often be decomposed into a set of parts, where a specific relational structure of parts determines the object label. Structured Output Learning has established as the branch of machine learning, addressing the learning of models of such structures. This thesis presents a novel Structured Output Learning approach, combining a graph-based and a probabilistic model representation. It learns from examples with varying numbers of attributed parts and relations, and assigns likelihoods to the existence of object structures. The learning of models is well founded in graph theory, and a comprehensive analysis of its properties is given. Different applications for learnt models are introduced, with a focus on classification of object structures, and recognition of object structures in a set of cluttered objects. A complete probabilistic framework for solving both tasks is presented. The proposed model representation and learning are data-independent, and can be applied to many different kinds of domains. To demonstrate the feasibility and the performance of the new model, experimental evaluation on a variety of four different benchmark datasets is presented, which includes applications in high-level image understanding, classification of toxic molecules, natural language processing and behaviour model learning.

Lingua: Inglese

Pubblicato: 2013