Descrição semântica de objetos em imagens baseada na Teoria dos Protótipos

• Ano de defesa: 2020
• Autor(a) principal: Omar Vidal Pino
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; Brasil; ICX - DEPARTAMENTO DE CIÊNCIA DA COMPUTAÇÃO; Programa de Pós-Graduação em Ciência da Computação; UFMG
• Palavras-chave em Português: Visão computacional; Aprendizagem profunda; Teoria dos Protótipos; Efeitos prototípicos; Descrição semântica; Computer vision; Deep learning; Prototype Theory; Prototypicality effects; Semantic description; Computação – Teses; Teoria dos protótipos – Teses; Aprendizado profundo – Teses; Visão computacional – Teses
• Link de acesso: http://hdl.handle.net/1843/34969; https://orcid.org/0000-0002-9917-3838

This research aims to build a model for semantic description of objects based on features detected in images. We introduce a novel semantic description approach inspired on the Prototype Theory foundations. Inspired by the human approach used for representing categories, we propose a novel Computational Prototype Model (CPM) that encodes and stores the central semantic meaning of the object’s category: the semantic prototype. Our CPM model is used to represent and construct the semantic prototypes of object categories using Convolutional Neural Networks (CNN). The proposed Prototype-based Description Model uses the CPM model to describe an object highlighting its most distinctive features within the category. Our Global Semantic Descriptor (GSDP) builds discriminative, low-dimensional and semantically interpretable signatures that encode the semantic information of the objects using the constructed semantic prototypes. Our semantic descriptor use the proposed Prototypical Similarity Layer (PS-Layer) to retrieves the category prototype using the principle of categorization based on prototypes. In our experiments, using publicly available datasets, we show that: i) the proposed CPM model adequately simulates the internal semantic structure of the categories; ii) the proposed semantic distance metric can be understood as the object typicality score within a category; iii) our semantic classification method based on prototypes can improve the performance and interpretation of CNN classification models; iv) our semantic descriptor encoding ignificantly outperforms others state-of-the-art image global encoding in clustering and classification tasks.