Representation learning and characterization of complex networks with applications in computer vision

Author(s)

Lucas Correia Ribas

• Document type: Doctoral Thesis
• Press: São Carlos.
• Institution: Universidade de São Paulo (USP). Instituto de Ciências Matemáticas e de Computação (ICMC/SB)
• Defense date: 2021-11-22

Examining board members

Odemir Martinez Bruno; Hélio Pedrini; Eraldo Ribeiro Junior; Agma Juci Machado Traina

• Advisor: Odemir Martinez Bruno

Abstract

Complex networks have been used as a study tool in several areas of science due to their multidisciplinary character and innovative perspective in relation to traditional data analysis. In computer vision, several approaches based on complex networks have been proposed over the last few years with promising results. Such success is explained by the ability to model and quantify the complexity present in many images, especially those arising from nature, which has a nonlinear behavior. Although the results are promising, optimized modeling techniques and more robust characterization methods to describe the complexity of the network topology are needed. This work aims to investigate and propose new modeling and characterization techniques for complex networks focusing on application to computer vision problems. Specifically, two classic problems of image and video analysis were investigated: texture (gray-level, color and dynamics) and shape analysis. Regarding modeling, an efficient and optimized approach to map texture images and videos into directed complex networks was studied. Regarding the characterization, based on classical metrics and cellular automata, two improvements were proposed: (i) binary patterns in network automata (LLNA-BP) for characterization of complex networks and (ii) distance transform networks for shape analysis. Furthermore, the learning of representations using randomized neural networks (RNNs) was investigated; a simple architecture and fast learning algorithm. In this sense, as a new form of characterization, representations formed by the weights of the output layer of RNNs trained with topological characteristics of complex networks were proposed. By combining the developed techniques of modeling and learning representations, several methods were proposed for analyzing gray-level, color and dynamic textures and shape analysis. Promising results have been achieved by the developed methods in comparison to literature methods in classification tasks using several benchmark datasets. Additionally, to assess the potential of the developed methods, five applications were investigated in real problems in the areas of biology, botany, physical-chemistry and medicine, achieving interesting results and contributing to the development of these areas. (AU)

• FAPESP's process: 16/23763-8 - Modeling and analysis of complex networks for computer vision
• Grantee: Lucas Correia Ribas
• Support Opportunities: Scholarships in Brazil - Doctorate