27 Jun 2025
09:00 Doctoral defense fully remotely
Topic on
Machine Learning Techniques for Image-Based Rock Classification
Student
Soroor Salavati
Advisor / Teacher
Anderson de Rezende Rocha - Co-advisor: Alexandre Mello Ferreira
Brief summary
Accurate classification of rock images is essential for reliable lithological analysis in geosciences. However, this task remains challenging due to subtle visual similarities between classes, data imbalance, and scarcity of annotations. This thesis presents a comprehensive, multi-stage framework that systematically evolves from traditional machine learning methods to advanced deep learning architectures to address these challenges. In the first stage, we establish a baseline using conventional classification models on heterogeneous rock datasets. Although simple and interpretable, these methods have limitations in capturing the complex textures and structures of geological samples. To overcome these limitations, we propose a hybrid cost function that combines categorical cross-entropy with the OHEM technique, which emphasizes difficult samples, reduces class imbalance, and improves the model’s ability to distinguish subtle features. Building on the learnings from the previous stages, we use a transformer-based architecture (DINOv2) for global feature extraction, combined with a CDC network to capture fine-grained local details. This architectural change represents a strategic advancement over the previous design. To increase class separability and model robustness, we introduce a joint cost function that combines Focal Loss and Center Loss—a refined combination that promotes intra-class compression and biases learning toward more difficult examples. Extensive experiments demonstrate that this improved model outperforms previous approaches and generalizes well on imbalanced geological datasets. To further improve performance, we incorporate a diffusion-based data augmentation strategy that synthetically expands the training data while preserving structural integrity. This final step strengthens the model’s generalization ability by addressing data sparsity during training. Overall, this thesis proposes a robust, end-to-end solution for rock image classification, integrating traditional and modern learning paradigms. The methodology has significant implications for reservoir characterization in the oil and gas industry and can be extended to other domains with similar data constraints, such as medical imaging or remote sensing.
Examination Board
Headlines:
| Anderson de Rezende Rocha | IC / UNICAMP |
| Diego Nunes Brandão | CEFET/RJ |
| George Darmiton da Cunha Cavalcanti | CIn / UFPE |
| Marcelo da Silva Reis | IC / UNICAMP |
| Levy Boccato | FEEC / UNICAMP |
Substitutes:
| Helena de Almeida Maia | IC / UNICAMP |
| Siovani Cintra Felipussi | DComp-So / UFSCar |
| João Paulo Papa | FC / UNESP |