Study and development of self-supervised representation learning techniques for image analysis: application in materials science

• Grant number: 25/14029-8
• Support Opportunities: Regular Research Grants
• Start date: February 01, 2026
• End date: January 31, 2029
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: Lucas Correia Ribas
• Grantee: Lucas Correia Ribas
• Host Institution: Instituto de Biociências, Letras e Ciências Exatas (IBILCE). Universidade Estadual Paulista (UNESP). Campus de São José do Rio Preto. São José do Rio Preto , SP, Brazil
• City of the host institution: São José do Rio Preto
• Associated researchers: JENNANE Rachid ; Osvaldo Novais de Oliveira Junior

Abstract

Image analysis, essential in various fields such as medicine and materials science, relies on visual representations to describe features such as texture, shape, and microscopic structure.Artificial neural network-based techniques have achieved remarkable results in computer vision, but still face important challenges, particularly in domains with limited labeled data, uncertain classes, and complex textural patterns, such as images of advanced materials, including sensors, biosensors, and nanomaterials.These materials, often used in critical applications such as medical diagnosis and environmental monitoring, typically exhibit high visual variability or intricate patterns, making their computational analysis an emerging and underexplored challenge.In this context, this project proposes to investigate and develop self-supervised learning techniques for image representation, aiming to overcome current limitations and improve tasks such as classification and diagnosis in materials science applications.To achieve this, the proposal is organized around three complementary research fronts that integratively explore the potential of Randomized Neural Networks (RNNets) and complex systems techniques in the context of self-supervised learning of visual representations.The main theoretical research directions include:(i) instance-based self-supervised learning using RNNets for fast processing and robustness in low data availability scenarios;(ii) self-supervised representation learning with large models, leveraging complex systems techniques to enrich learned features through pretext tasks, masking strategies, and the generation of complementary views that highlight the complexity of microtextural structures and morphological patterns relevant to the applications;(iii) auto-encoding of textural features from backbones (exploring new ViT and CNN architectures) using RNNets, aiming to optimize the extraction and representation of features relevant to specific tasks.These approaches will be applied to images of advanced materials, such as sensors and biosensors, with the goal of developing new tools for tasks including diagnosis, structural characterization, monitoring of physicochemical properties, and automated detection of relevant patterns.In particular, the computational analysis of these images, provided by collaborators, remains relatively unexplored, representing a promising area for investigation in conjunction with machine learning techniques.This project ultimately aims to foster a synergy between theory and application, where practical challenges serve as catalysts for new methodologies, while theoretical advances lead to effective computational tools for automated material analysis, contributing to progress in both materials science and machine learning. (AU)