TBD-Y: Automatic tea bud detection with synergistic object-spatial attention and global-local attention guided feature fusion
Automatic Tea Bud Detection (TBD) is a critical technology in intelligent tea-picking systems. Nevertheless, challenges, such as complex environments and the high visual similarity between tea buds and backgrounds, frequently result in false detection and missed detection, especially for small tea b...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-12-01
|
| Series: | Smart Agricultural Technology |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S2772375525002990 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Automatic Tea Bud Detection (TBD) is a critical technology in intelligent tea-picking systems. Nevertheless, challenges, such as complex environments and the high visual similarity between tea buds and backgrounds, frequently result in false detection and missed detection, especially for small tea buds. To address these issues, in this paper, an automatic TBD method is proposed, which is built upon the YOLOv11 object detection framework, named TBD-Y. Firstly, a Synergistic Object-Spatial Attention (SOSA) mechanism is proposed, which incorporates the proposed Local Context Attention (LCA) mechanism to enhance the features in both spatial and regional dimensions. It enables the network to focus more on the tea bud regions, and suppress the interference from background noise. Secondly, a Global-local Attention Guided Feature Fusion (GAGFF) strategy is designed. It consists of two branches: one branch enhances low-resolution, high-level features containing rich global semantic information, while the other branch strengthens high-resolution features that preserve low-level visual details. The fusion of these two branches improves the representation capability of the features. The SOSA and GAGFF are integrated into the YOLOv11 framework, constructing three variants of the TBD model with different parameter scales, named TBD-Y-L, TBD-Y-M, and TBD-Y-S. Experimental results on the self-built TBD dataset and the publicly available Global Wheat Head Dataset 2021 (GWHD_2021) demonstrate that the proposed TBD-Y-L outperforms the existing methods, achieving superior detection accuracy. Furthermore, the TBD-Y-S model exhibits improved detection accuracy compared to YOLOv11-L, while maintaining lower model parameters and computational complexity. |
|---|---|
| ISSN: | 2772-3755 |