16Jul2026
09:00 Doctoral defense Room 85 of IC2
Topic on
Diversity-Informed Deep Learning Selection and Model Ensembles for Common Bean Disease Detection
Student
Rubens de Castro Pereira
Advisor / Teacher
Hélio Pedrini - Co-supervisors: Murillo Lobo Junior and Díbio Leandro Borges
Brief summary
Object detection in deep learning has advanced significantly in architectures and methodologies, but remains challenging in real-world field images due to small objects, occlusions, background noise, class imbalance, and limited annotations. In agriculture, this technique supports decision-making, such as the detection of pests, insects, and plant diseases. White mold, caused by the fungus Sclerotinia sclerotiorum, is especially challenging due to its wide host range, its persistence in the soil, and its difficult management, affecting more than 400 plant species, including the common bean (Phaseolus vulgaris L.). Current methods typically utilize individual detectors or combinations of multiple detectors with distinct architectures and performance characteristics. Although late-fusion combination methods achieve high detection performance, they require considerable computational and training costs. Diversity-based detector selection emerges as a promising strategy for combining detectors, reducing the use of computational resources and improving detection performance. This thesis investigated the use of detector diversity to select the most promising detectors for late-fusion combination of predictions, aiming to improve detection. Two distinct approaches have been proposed for the selection of detectors, based on diversity and effectiveness, for the subsequent combination of predictions. The Diversity-Informed Object Detector Fusion (DiODeFusion) framework was introduced, which provides a list of detectors selected based on diversity measures between pairs of detectors, for combination by late fusion of predictions. The second framework, Diversity-Informed Clustering for Object Detector Fusion (DiCODeFusion), selects the most promising detectors through hierarchical clustering based on diversity and performs the fusion of predictions. In addition, a new set of expert-annotated images of the fungus S. was developed. Sclerotinia and White Mold (SWM) is a study of sclerotiorum and white mold from real images of common bean fields. The reference detectors evaluated in the SWM dataset include SSD, Faster R-CNN, YOLOv8, YOLOv9, YOLOv10, DETR, and adapted TransUNet. DiODeFusion was evaluated exclusively on the SWM dataset, while DiCODeFusion considered different datasets including SWM, NC, FDWE, and COCO 2017. The reference detectors were trained and evaluated separately on their respective datasets, according to the experimental protocols. The DiODeFusion results revealed that detector selection based on diversity and late fusion outperformed the best individual detector in terms of F1-score, even though this framework requires a parameter (k) to select the top-k detectors. Furthermore, the DiCODeFusion results improved detection performance by selecting a diverse and effective cluster of detectors from the hierarchy for prediction fusion. The contributions to object detection were presented, along with advances in deep learning model combination strategies, achieved through the careful selection of object detection models for late fusion. In summary, this thesis deepens the understanding of the diversity in detector combination approaches.
Examination Board
Headlines:
| Hélio Pedrini | IC / UNICAMP |
| Jayme Garcia Arnal Barbedo | EMBRAPA |
| Edimilson Batista dos Santos | DCC / UFSJ |
| Juliana Aparecida Fracarolli | FEAGRI / UNICAMP |
| Washington Luiz Miranda da Cunha | IC / UNICAMP |
Substitutes:
| Helena de Almeida Maia | IC / UNICAMP |
| Ronaldo Cristiano Prati | CMCC / UFABC |
| Moacir Antonelli Ponti | ICMC / USP |