MSIFD-Net: tea stem recognition network based on multi-scale interaction combined feature decoupling
Year: 2026
Type: article
Abstract: Accurate tea bud detection is essential for intelligent tea-picking robots. After multiple harvests, tea stems appear more frequently, and many tea bud detection models misclassify stems as buds due to subtle differences in shape, texture, and color. Such errors reduce detection accuracy and may damage the end effector when rigid stems are mistakenly selected. In order to reduce the false identification of tea stems as buds, this study proposes a tea stem recognition network based on multi-scale interaction combined with feature decoupling (MSIFD-Net), which operates at the output of a tea bud detection model. The network exploits shallow and deep features to enhance fine-grained discrimination. An Adaptive Channel–Spatial Attention module suppresses irrelevant responses in multi-scale features and emphasizes key spatial regions and channels. A multi-scale interaction structure extracts complementary information from global features across different scales. In addition, a feature decoupling module separates feature components into independent subspaces to reduce redundancy. On the ZC108 dataset, the integration of MSIFD-Net into YOLOv11 achieved a peak AP of 91.28%, improving by 10.23 percentage points compared with YOLOv11 without MSIFD-Net. Furthermore, the FDR decreased markedly from 22.22% to 1.1%, representing a reduction of 21.12 percentage points. These results demonstrated that the proposed method effectively reduced false detections caused by residual tea stems and decreased the risk of unnecessary picking operations and potential damage to the tea-picking robot end effector.
Institution: Zhejiang Sci-Tech University
Language: English
Topic: Smart Agriculture and AI
Subfield: Plant Science
Domain: Life Sciences
Open Access status: gold
Estimated APC paid: $323
OpenAlex ID: W7225825385