Kernel Estimation in FLIM-based CNNs for Object Detection

• Grant number: 24/08332-7
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date: August 01, 2024
• End date: October 31, 2025
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: Alexandre Xavier Falcão
• Grantee: João Deltregia Martinelli
• Host Institution: Instituto de Computação (IC). Universidade Estadual de Campinas (UNICAMP). Campinas , SP, Brazil
• Associated scholarship(s): 24/23772-3 - Parasite Egg Segmentation With a FLIM-BoFP Network, BE.EP.IC

Abstract

In recent years, Convolutional Neural Networks (CNNs) have seen immense success in creating models for decision-making and decision-support systems. However, as their complexity increases in attempts to solve harder problems, heavyweight CNNs pose challenges in data requirements, interpretability, and computational resources. Feature Learning from Image Markers (FLIM), a recent methodology to estimate CNN kernels from image markers without backpropagation, has shown promising results in addressing these issues. By leveraging user-drawn markers on relevant regions of a few training images and clustering image patches extracted from those markers for filter (kernel) estimation, the network requires considerably less annotated data, producing results easily interpretable by the network designer. This undergraduate research project aims to conduct a comparative analysis of two main variations of kernel estimation for FLIM networks for object detection. While their authors have informally explored these variations, a systematic comparison is necessary to understand their relative strengths and weaknesses. The comparative analysis will focus on critical metrics such as model accuracy, size, and computational requirements. This analysis will provide insights into the strengths and weaknesses of each method, facilitating decisions regarding their applicability in resource-constrained environments. This research seeks to advance the design of FLIM-based CNNs by elucidating the trade-offs between kernel estimation approaches. One of the kernel estimation methods has not been published yet, so the findings in this study have potential to present it in a scientific paper. The project also considers an overseas internship with one of the supervisor's collaborators, Prof. Jeferson dos Santos, at the University of Sheffield.