28Jul2025
09:00 Doctoral defense room 85 of IC2
Topic on
Building Flyweight CNNs for Salient Object Detection
Student
Leonardo de Mélo João
Advisor / Teacher
Alexandre Xavier Falcao - Co-supervisors: Ewa Kijak and Jancarlo Ferreira Gomes
Brief summary
Salient Object Detection (SOD) often relies on deep neural networks that are computationally intensive, require expensive training, and large annotated datasets—which poses challenges in biomedical image analysis, where expert annotation is difficult. Furthermore, interpretability is crucial for automated diagnostics. This PhD thesis addresses these issues by developing \textbf{compact and explainable SOD models} such as Convolutional Neural Networks (CNNs) that require few annotations and low computational power. The hypothesis is that lightweight models can achieve results equivalent to state-of-the-art complex models when tuned for specific applications with limited labeled data. To achieve this goal, we leveraged \textbf{domain-specific knowledge} within the \textbf{Feature Learning by Image Marker (FLIM)} framework, which allows CNNs to learn from discriminative regions of representative images indicated by experts. A significant contribution is the introduction of \textbf{unsupervised adaptive decoders}, eliminating the need for annotated data and backpropagation. These decoders estimate feature importance dynamically, aiding visualization during training. A \textbf{new formulation of FLIM} also separates the marker sets used in filter learning and Marker-Based Normalization (MBN), enabling deeper analysis of the effects of MBN. \textbf{marker bots} and superpixel-extended markers were introduced to control sampling rates. \textbf{separable kernels with multi-dilation} were implemented without backpropagation to further optimize efficiency, allowing single-layer adaptive decoders to capture information at multiple scales. A \textbf{pruning algorithm} specific to FLIM was also developed to eliminate redundant kernels. Validation on biomedical datasets—Schistosoma mansoni eggs on optical microscopy and brain tumors on MRI—showed that flyweight CNNs outperform standard FLIM networks when competing with heavyweight models in data-sparse scenarios. This work establishes a foundation for saliency-based object detection as a preliminary task for object identification, and extends the flyweight learning methodology for future investigations in classification tasks.
Examination Board
Headlines:
| Alexandre Xavier Falcão | IC / UNICAMP |
| David Menotti Gomes | DInf / UFPR |
| Laurent Najman | ESIEE Paris |
| Marcelo da Silva Reis | IC / UNICAMP |
| Wallace Correa de Oliveira Casaca | IBILCE/UNESP |
Substitutes:
| Marcos Medeiros Raimundo | IC / UNICAMP |
| Nina Sumiko Tomita Hirata | IME / USP |
| Luciano Rebouças de Oliveira | DCC/UFBA |