Examinando por Materia "Deep learning"

Mostrando 1 - 1 de 1

Detection of Solanum betaceum Cav fruit maturity using YOLO11-based deep learning
(Frontiers Media S.A., 2026-09-10) Quiñones Huatangari, Lenin; Vásquez Pérez, Héctor Vladimir; Valqui Valqui, Lamberto; Palomino Ojeda, Jose Manuel; Saravia Navarro, David; Bardales Escalante, William; Silva Valqui, Gelver; Chavez Jalk, Antony; Cruz Caro, Omer; Valqui, Leandro
The tree tomato (Solanum betaceum Cav.) is a Solanaceae fruit native to South America with nutritional, functional, and economic value. However, assessing fruit maturity and making harvest-related decisions continue to rely primarily on manual visual inspection, which can be subjective and inconsistent under field conditions. The objective of this study was to evaluate the performance of YOLO11-based object detection models for the automatic detection of the ripeness of Solanum betaceum fruits in a real-world agricultural setting. To this end, a dataset of 200 images captured with smartphones was used, in which 1,361 fruits were annotated with bounding boxes corresponding to green and ripe fruits. The YOLO11n, YOLO11s, and YOLO11m models were trained with an input resolution of 640. For each setting, cross-validation (k = 5) was used, and data augmentation was applied to the training subsets. YOLO11m achieved the best overall performance, as it combined high values for precision, recall, and mAP. On the test set, YOLO11m achieved a precision of 0.9878 ± 0.0055, a recall of 0.9750 ± 0.148, and an mAP50 of 0.9618 ± 0.0188 for green fruit, as well as a precision of 0.9950 ± 0.0041, a recall of 0.9958 ± 0.0039, and an mAP50 of 0.9920 ± 0.0047 for ripe fruits. YOLO11s also demonstrated competitive performance, particularly in the detection of ripe fruits, with an mAP50 of 0.9894 ± 0.0070 and an mAP50–95 of 0.9754 ± 0.0081, positioning it as an efficient alternative given its substantially lower computational cost. In contrast, YOLO11n, the lightest architecture, showed the lowest performance for green fruits (mAP50-95: 0.8557 ± 0.0611), reflecting the limitations of its representational capacity.