- Article
Tea sprout detection is critical for intelligent tea harvesting. However, existing recognition methods still face challenges, including dense target distribution, limited feature extraction, and high false detection rates. To address these issues, this study establishes a multi-scene, multi-variety image database and proposes TeaP2AT-SDI+, an enhanced YOLOv11n model with three targeted improvements: a P2 high-resolution detection head is introduced to recover fine-grained spatial details of small tea sprouts; a partial attention convolution (PATConv) module is adopted to improve feature representation efficiency; and a semantic and detail injection (SDI) module with scale-aware gating is designed to stabilize cross-scale semantic propagation. Experimental results demonstrate that the improved model achieves an mAP@0.5 of 91%, a recall of 85.9%, and a precision of 84.2%, outperforming the baseline YOLOv11n by 10.6, 10.2, and 9.9 percentage points, respectively. TeaP2AT-SDI+ also consistently outperforms YOLOv5n, YOLOv8n, YOLOv26n, and R-DETR in overall detection accuracy. Notably, it requires only 4.34 M parameters, which is 8.7× fewer than R-DETR, and achieves a latency of 2.9 ms. For object classes, the model attains an mAP@0.5 of 95.4% for “Bud with leaves” and 86.5% for “One bud”, effectively supporting the multi-grade classification aligned with common Chinese tea-picking grades. It demonstrates robust adaptability across diverse lighting and background conditions, providing reliable technical support for real-time tea sprout detection and intelligent tea harvesting automation.
14 November 2026










