Ingeniería Electrónica
YOLO vs U-Net en detección y clasificación de granos de café según su madurez
This work presents a comparative study between the YOLO and U-Net architectures for the detection, segmentation, and classification of coffee fruit maturity stages using computer vision and deep learning techniques. The main objective is to analyze the behavior of both architectures using the same image dataset, evaluating their capabilities to identify coffee fruits at different maturity stages under real agricultural conditions. To achieve this, an experimental workflow was developed including multiple training configurations, data augmentation techniques, and class balancing strategies designed to mitigate the class imbalance present in the dataset. For YOLO, different transfer learning and progressive balancing strategies were evaluated, while for U-Net, multiclass configurations with pretrained encoders, targeted data augmentation, and weighted sampling mechanisms were implemented. The obtained results demonstrate that YOLO achieves better performance in object detection and individual fruit counting tasks, whereas U-Net provides a more detailed spatial representation of the regions associated with each maturity stage. Additionally, the study shows that balancing and data augmentation strategies significantly improve performance on minority classes. Finally, this research establishes the advantages, limitations, and potential applications of each architecture within the agricultural domain, providing an experimental foundation for future automated monitoring and coffee maturity analysis systems.