21Jul2025
09:00 Master's Defense room 85 of IC2
Topic on
Integrating Label-Based Filter Learning into Images with Cellular Automatons
Student
Felipe Crispim da Rocha Salvagnini
Advisor / Teacher
Alexandre Xavier Falcao
Brief summary
Deep learning approaches typically require large annotated datasets and increasingly complex architectures. This paradigm presents significant challenges in resource-constrained environments, particularly for medical applications in developing countries, where data annotation is costly and computational resources are limited. Furthermore, many real-world problems, such as parasite egg detection, involve fewer categories than general computer vision tasks, suggesting that simplified approaches may be feasible. Therefore, this master’s thesis examines the Feature Learning from Image Markers (FLIM) methodology, which allows domain experts to design convolutional encoders directly from image patches, providing control over the encoder complexity. By integrating an adaptive decoder with a FLIM encoder, we create models that eliminate the need for backpropagation and substantially reduce annotation requirements (typically to only 3-4 images). Combining FLIM networks with Cellular Automata (CA) creates a complete pipeline. CA serves as a post-processing technique; additionally, FLIM facilitates CA initialization by leveraging user knowledge without requiring interaction for each image. We evaluate FLIM-CA integration for salient object detection on two medical datasets: parasite egg detection and brain tumor detection. Our results demonstrate competitive performance compared to convolutional deep learning methods, with improvements of 13% to 20% in metrics such as F-Score and uWF, using only a fraction of the parameters (thousands vs. millions). Our work proposes a multi-level FLIM-CA system that builds on the hierarchical capabilities of FLIM encoders.
Examination Board
Headlines:
| Alexandre Xavier Falcão | IC / UNICAMP |
| Fátima de Lourdes dos Santos Nunes Marques | EACH / USP |
| João Paulo Papa | FC / UNESP |
Substitutes:
| Marcos Medeiros Raimundo | IC / UNICAMP |
| Nina Sumiko Tomita Hirata | IME / USP |