Full text
| |
| Author(s): |
Rafael de Oliveira Werneck
Total Authors: 1
|
| Document type: | Doctoral Thesis |
| Press: | Campinas, SP. |
| Institution: | Universidade Estadual de Campinas (UNICAMP). Instituto de Computação |
| Defense date: | 2019-05-31 |
| Examining board members: |
Ricardo da Silva Torres;
Luciano Rebouças de Oliveira;
David Menotti;
Alexandre Mello Ferreira;
Fábio Luiz Usberti
|
| Advisor: | Ricardo da Silva Torres |
| Abstract | |
Many real-world situations can be modeled through objects and their relationships, like the roads connecting cities in a map. Graph is a concept derived from the abstraction of these situations. Graphs are a powerful structural representation, which encodes relationship among objects and among their components into a single formalism. This representation is so powerful that it is applied to a wide range of applications, ranging from bioinformatics to social networks. Thus, several pattern recognition problems are modeled to use graph-based representations. In classification problems, the relationships among objects or among their components are exploited to achieve effective and/or efficient solutions. In this thesis, we investigate the use of graphs in classification problems. Two research venues are followed: 1) proposal of graph-based multimodal object representations; and 2) proposal of learning-based approaches to support graph matching. Firstly, we investigated the use of the recently proposed Bag-of-Visual-Graphs method in the representation of regions in a remote sensing classification problem, considering the spatial distribution of interest points within the image. When we combined color and texture representations, we obtained effective results in two datasets of the literature (Monte Santo and Campinas). Secondly, we proposed two new extensions of the Bag-of-Graphs method to the representation of multimodal objects. By using these approaches, we can combine complementary views of different modalities (e.g., visual and textual descriptions). We validated the use of these approaches in the flooding detection problem proposed by the MediaEval initiative, achieving 86.9\% of accuracy at the Precision@50. We addressed the graph matching problem by proposing an original framework to learn the cost function in a graph edit distance method. We also presented a couple of formulations using open-set recognition methods and complex network measurements to characterize local graph properties. To the best of our knowledge, we were the first to conduct the cost learning process as an open-set recognition problem and to exploit complex network measurements in such problems. We have achieved effective results, which are comparable to several baselines in graph classification problems (AU) | |
| FAPESP's process: | 16/18429-1 - A bag-of-graphs approach for cross-modal representations |
| Grantee: | Rafael de Oliveira Werneck |
| Support Opportunities: | Scholarships in Brazil - Doctorate |