Copertina

LettuCeV: sistema de visão computacional para detecção de doenças em raízes de alface hidropônica

Salvato in:
Dettagli Bibliografici
Autore principale: Agostini, Pedro Lucas Trentin
Data di pubblicazione: 2026
Natura: Master thesis
Lingua: por
Fonte: Repositório Institucional da UPF
Download full: https://repositorio.upf.br/handle/123456789/10250
Riassunto: The identification of diseases in the roots of hydroponically grown lettuce is relevant for phytosanitary management in greenhouses; however, it is challenging to standardize in daily practice due to its reliance on manual visual inspection and variable image acquisition conditions. This work presents LettuCeV, a computer vision system for disease detection in hydroponic lettuce roots. The system integrates a YOLOv8-seg model into an application designed for image acquisition and processing, cloud-based inference, and data recording. The methodology included the construction of a dedicated dataset comprising 651 images of curly lettuce roots, totaling 772 annotations, of which 331 correspond to diseased roots and 441 to healthy roots. The trained model achieved satisfactory performance in root condition classification (F1-Score ≈0,7902) and segmentation (Dice ≈0,7770), with particular emphasis on detecting healthy roots (Accuracy ≈0,93). The system was evaluated through a pilot study involving 40 images captured by 11 volunteers in controlled experimental greenhouses, to assess the model’s computational performance and the application’s usability under nearreal-world conditions. For classification tasks, the model maintained performance metrics comparable to those observed during training, indicating its potential to support practical plant screening activities and historical data recording. Conversely, segmentation evaluation based on pixel-level comparison between predictions and manual annotations revealed difficulties in finely delineating regions with subtle lesions (Dice ≈0,132), particularly regions presenting early-stage disease. The application demonstrated high user acceptance, achieving an excellent overall usability score (SUS = 87.05). In practical use, inference results were returned via API ≈ 15s after image submission for analysis, while the complete user-perceived process required ≈ 2min per record. In summary, this study demonstrated the technical feasibility and application potential of LettuCeV as a decision-support tool and identified clear directions for future model improvements.
Riassunto:The identification of diseases in the roots of hydroponically grown lettuce is relevant for phytosanitary management in greenhouses; however, it is challenging to standardize in daily practice due to its reliance on manual visual inspection and variable image acquisition conditions. This work presents LettuCeV, a computer vision system for disease detection in hydroponic lettuce roots. The system integrates a YOLOv8-seg model into an application designed for image acquisition and processing, cloud-based inference, and data recording. The methodology included the construction of a dedicated dataset comprising 651 images of curly lettuce roots, totaling 772 annotations, of which 331 correspond to diseased roots and 441 to healthy roots. The trained model achieved satisfactory performance in root condition classification (F1-Score ≈0,7902) and segmentation (Dice ≈0,7770), with particular emphasis on detecting healthy roots (Accuracy ≈0,93). The system was evaluated through a pilot study involving 40 images captured by 11 volunteers in controlled experimental greenhouses, to assess the model’s computational performance and the application’s usability under nearreal-world conditions. For classification tasks, the model maintained performance metrics comparable to those observed during training, indicating its potential to support practical plant screening activities and historical data recording. Conversely, segmentation evaluation based on pixel-level comparison between predictions and manual annotations revealed difficulties in finely delineating regions with subtle lesions (Dice ≈0,132), particularly regions presenting early-stage disease. The application demonstrated high user acceptance, achieving an excellent overall usability score (SUS = 87.05). In practical use, inference results were returned via API ≈ 15s after image submission for analysis, while the complete user-perceived process required ≈ 2min per record. In summary, this study demonstrated the technical feasibility and application potential of LettuCeV as a decision-support tool and identified clear directions for future model improvements.

Documenti analoghi