Tea bud detection in natural environments based on YOLOv11n-WELA

• YOLOv11n-WELA constructs a Weighted Fusion-based Efficient Multi-Scale Feature Pyramid Network (WF-EMSFPN), which enhances the feature representation ability of the network through a dynamic weight assignment algorithm for cross-layer feature fusion. • LADH structure in YOLOv11n-WELA reduces computational complexity, suitable for mobile use. • The ADown module in YOLOv11n-WELA reduces the loss of fine-grained features and enhances the model’s detection capability. • YOLOv11n-WELA significantly reduces model size and computational complexity while maintaining detection accuracy, making it suitable for real-time detection tasks in complex tea plantation scenarios. This advancement provides a high-precision lightweight visual perception system for intelligent harvesting robotics. • YOLOv11n-WELA model achieves stable real-time detection performance on the Jetson Nano edge deployment platform, with a frame rate consistently maintained at 31 FPS. This validates the model's feasibility for practical deployment on lightweight devices. Aiming at the issue of reduced detection precision in models resulted from the small target size, dense growth, and high color similarity to mature leaves of tea buds, along with computational constraints and storage bottlenecks for the deployment of deep models on the mobile platform, this study proposes a YOLOv11n-WELA detection algorithm based on multi-scale feature fusion and lightweight design. The algorithm uses YOLOv11n as the baseline model, and introduces lightweight ADown modules in the backbone network to effectively extract both local and global features of tea buds. Furthermore, a Weighted Fusion-based Efficient Multi-Scale Feature Pyramid Network (WF-EMSFPN) is innovatively designed, establishing a cross-scale feature weighted fusion mechanism to effectively integrate hierarchical feature information. Additionally, the detection head adopts LADH structure, which effectively reduces the model’s parameters and storage space. The results indicate that YOLOv11n-WELA obtains 90.67% mAP@0.5, an increase of 0.91% over the baseline model, while reducing its model size, floating-point operations (FLOPs), and parameters by 32.73%, 25.40%, and 37.98%, respectively. Through ablation experiments and multi-dimensional comparison experiments, the algorithm’s tea bud detection ability is confirmed. The improved model was further tested on an NVIDIA Jetson Nano edge device, demonstrating stable real-time performance with a frame rate of 31 FPS. This meets the requirements of practical applications, verifies deployment feasibility, and provides effective support for the implementation of intelligent agricultural technologies.