|
Title: |
|
Authors:
|
|
Abstract: Accurate tea-bud detection is essential
for intelligent tea harvesting, yield estimation, and precision tea production.
However, practical deployment remains challenging because of dense bud
distributions, severe occlusion, variable illumination, and the high cost of
large-scale manual annotation. To address these challenges, we propose
TeaBud-YOLO11++, an annotation-efficient semi-supervised framework that extends
YOLO11 through teacher–student learning and a novel Density-Aware Confidence
Recalibration (DACR) mechanism that adaptively recalibrates pseudo-label
confidence according to local bud density. By improving pseudo-label
reliability in densely populated scenes, the proposed framework enables more
effective use of unlabeled data while reducing annotation dependency. Experiments
on a field dataset of 1,500 images containing 68,895 manually annotated tea
buds showed that TeaBud-YOLO11++ achieved 93.33% precision, 91.21% recall,
95.08% mAP@0.5, and 75.45% mAP@0.5:0.95, outperforming representative YOLO11
variants, RT-DETR-L, and Faster R-CNN. Furthermore, DACR reduced the Expected
Calibration Error by 28.8%, Furthermore, DACR reduced the Expected Calibration
Error by 28.8%, while environmental robustness, mixed-domain evaluation, and
repeated-training analyses provided additional evidence of the framework's
effectiveness under diverse plantation conditions. These results demonstrate
that TeaBud-YOLO11++ provides a robust, annotation-efficient machine vision
solution for intelligent tea harvesting and precision tea production. DOI: http://dx.doi.org/10.51505/ijaemr.2026.11601 |
|
PDF Download |