Robust detection of dense small tea shoots across cultivars under occlusion and bud–leaf similarity for intelligent selective harvesting

Accurate and robust recognition of dense small tea shoots in complex fields is critical for automated picking, but occlusion and bud–leaf similarity degrade detection accuracy. To address these challenges, the YOLOv7-LEES model is proposed. The backbone integrates Efficient Channel Attention (ECA) to improve recall on dense small objects. The Neck network incorporates the Explicit Visual Centre (EVC) to capture complementary global and local features, increasing sensitivity and reducing misclassification under occlusion, while RepNet Cross-Stage Partial Attention ELAN4 Network (RepNCSPELAN4) reduces computation cost and parameters. SCYLLA-IoU (SIoU) loss with angle cost reduces localisation uncertainty and speeds convergence. YOLOv7-LEES achieves 92.0 % precision, 82.2 % recall, 90.3 % mAP0.5, and 88.0 % F1 score, with 35.5 M parameters and 116.3 FPS. Ablation studies show that EVC increases mAP0.5SS by 9.8 %, ECA enhances Rall by 2.3 %, and SIoU improves PSS by 4.9 %, while RepNCSPELAN4 reduces parameters by 12.9 % and computational costs by 8.3 %. Compared to the classical two-stage detectors (Faster R-CNN) and the most recent single-stage detector (YOLOv11m and YOLOv12m), YOLOv7-LEES offers advantages in both detection accuracy and speed. On a multi-cultivar dataset, YOLOv7-LEES achieves 93.35 % mAP0.5 and demonstrates consistent performance across varying weather conditions. It robust to bud overlap, occlusion by old leaves, and similarity between buds and leaves, demonstrating adaptability from sparse to dense shoot distributions, achieving 82.0 % mAP0.5 for small and 93.9 % mAP0.5 for large quantities. YOLOv7-LEES achieves robust dense tea-shoot detection under field conditions, supporting precise and efficient tea-picking automation.