Active learning for data-efficient segmentation of tea buds and leaves

Accurate recognition and segmentation of tea buds and leaves is critical for precision tea management. However, developing robust vision models for tea bud and leaf segmentation typically requires large volumes of manually annotated data, which are costly and time-consuming to obtain. To address this challenge, we propose an active learning framework for data-efficient segmentation of tea buds and leaves. The proposed framework adopts an iterative training strategy that begins with a small set of labeled images and progressively improves the segmentation model by selectively annotating the most informative unlabeled samples. These samples are identified using mean confidence scores from a YOLO-based segmentation network, prioritizing uncertain instances to maximize learning efficiency. The proposed approach is evaluated using multi-view image datasets collected from tea plantations, including top-down and 45° perspectives, to investigate the influence of camera viewpoint on segmentation performance. In addition, segmentation accuracy is assessed across different tea plucking standards, including single buds, one-bud-one-leaf, and one-bud-two-leaves. Experiments show that, under the same annotation budget, the proposed framework achieves more than 10% higher accuracy compared to random sample selection. The performance gains are even more pronounced for 45° imaging, where our method outperforms random annotation by 14.7% in mAP@0.5 overall and by 19.0% specifically for bud detection.