Full text
| |
| Author(s): |
Renato Gomes Marcacini
Total Authors: 1
|
| Document type: | Master's Dissertation |
| Press: | São Carlos. |
| Institution: | Universidade de São Paulo (USP). Instituto de Ciências Matemáticas e de Computação (ICMC/SB) |
| Defense date: | 2023-12-21 |
| Examining board members: |
Agma Juci Machado Traina;
João do Espírito Santo Batista Neto;
Marcela Xavier Ribeiro;
Lúcio Fernandes Dutra Santos
|
| Advisor: | Agma Juci Machado Traina |
| Abstract | |
Indexing and retrieving information organized in databases efficiently and accurately to respond to queries requested by users and specialists are pillars of the engineering and data science process. In the medical context, complex image-type datas management and semantic extraction are fundamental for decision-making. This type of management requires extracting the essence (features) of these data and usually carrying out similar queries on such features, instead of using the complex data itself. The Metric Access Methods (MAMs) were developed as fundamental tools to process similarity queries supporting database management systems efficiently. MAMs use fixed distance functions to build the metric tree, which in turn prevents a MAM from being able to index elements using two or more distance functions in the same index. In many cases, the use of traditional distance functions fails to provide answers compatible with users perception of distance. This Masters level research developed novel mechanisms to deal with this issue, developing an approach that allows including weighted distance functions for processing similarity queries in MAM. A properly learned vector of weights allows weighting distance functions and improving data semantics, enhancing the query processing accuracy. For this intent, resources of Relevance Feedback (RF) were used, to capture the users perception regarding the similarity between images. This work proposes two methods to enhance the feature space and include weighted distance functions in the Slim-Tree MAM. The method Fusion Relevance Feedback (FRF) applies a pre-processing step combining traditional feature extractors and improving the feature space using RF, achieving equivalent and superior accuracies compared to deep learning techniques. The Tuning Metrics Relevance Feedback (TMRF) method infers weight vectors in the Slim-Tree and presents a reindexing methodology that keeps the structure optimized with the metric space improving. The analyses demonstrated that the TMRF method improves the metric space of the data set and keeps the MAM efficient, being 70% faster in relation to sequential strategies, with expressive gains in accuracy of up to 42% through learning by RF. (AU) | |
| FAPESP's process: | 21/00366-1 - Development of a metric access method with perceptual distance tuning to support similarity queries of medical images |
| Grantee: | Renato Gomes Marcacini |
| Support Opportunities: | Scholarships in Brazil - Master |