Enhancing FLIM Networks through Encoder Optimization and Learnable Graph Methods.

• Grant number: 25/19856-0
• Support Opportunities: Scholarships in Brazil - Doctorate
• Start date: March 01, 2026
• End date: September 30, 2029
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: Alexandre Xavier Falcão
• Grantee: Gilson Júnior Soares
• Host Institution: Instituto de Computação (IC). Universidade Estadual de Campinas (UNICAMP). Campinas , SP, Brazil

Abstract

Deep Learning-based Convolutional Neural Networks (CNN) have presented impressive results in different tasks related to computer vision and image processing. However, their need for large datasets and costly backpropagation training remains challenging in resource-constrained environments. Feature Learning From Image Markers (FLIM) was presented as a methodology for creating lightweight CNN encoders with minimal human effort in data annotation. This proposal builds on the applicant's MSc work on adaptive decoders for Salient Object Detection (SOD) using FLIM networks. The aim is to improve FLIM networks by: (I)creating adaptive decoders for other applications (e.g., extending them for multiclass problems); (II) investigating the adaptive decoders as an attention mechanism to enhance FLIM encoders; (III) exploiting the fast-training of FLIM encoders for Network Architecture Search (NAS); and (IV) creating methods to train the arc weights of a graph algorithm (e.g., for object delineation) as the head of a FLIM network. Adaptive decoders are a recent breakthrough that allows training the entire FLIM network without backpropagation (i.e., with no need for pixel-wise annotated images). By doing so, we aim to integrate parameter learning into graph-based image processing and extend neural networks to have more complex image operators. The methods will be validated on topics with which the advisor has extensive experience, such as medical and remote sensing image analysis. Some segments of the project will be developed as part of a BEPE project, which will provide the candidate with international experience. We are considering traditional partners of the advisor, such as Professors Laurent Najman (Gustave Eiffel University, ESIEE-Paris) and Devis Tuia (EPFL, Switzerland). We are also considering a recent opportunity for a joint PhD with Professor Xiaoyi Jiang at the University of Münster. (AU)