YOLO-LMTB: A Lightweight Detection Model for Multi-Scale Tea Buds in Agriculture

Area of Science:

• Computer Vision
• Machine Learning
• Agricultural Technology

Background:

• Tea bud detection is challenging due to complex environments with multi-scale variations, high density, and background color resemblance.
• Accurate and rapid detection of tea buds is crucial for automated tea-picking systems.

Purpose of the Study:

• To develop a lightweight, multi-scale object detection model for tea buds.
• To enhance the accuracy and efficiency of tea bud detection in complex agricultural settings.
• To provide a technical foundation for intelligent tea-picking systems.

Main Methods:

• Proposed YOLO-LMTB model based on YOLOv11n architecture.
• Introduced Multi-scale Edge-Refinement Context Aggregator (MERCA) module for improved feature perception.
• Developed Dynamic Hyperbolic Token Statistics Transformer (DHTST) module to enhance discriminative features and suppress background noise.
• Implemented Bidirectional Feature Pyramid Network (BiFPN) for adaptive feature fusion and reduced computational complexity.

Main Results:

• YOLO-LMTB achieved a 2.9% improvement in precision (P), 1.6% in mAP50, and 2.0% in mAP50-95 compared to the original model.
• Reduced model parameters by 28.3% and model size by 22.6%.
• Demonstrated strong generalization capabilities on public datasets, with each enhancement module boosting performance.

Conclusions:

• The YOLO-LMTB model effectively addresses the challenges of multi-scale tea bud detection in complex environments.
• The proposed MERCA, DHTST, and BiFPN modules significantly improve detection accuracy and efficiency.
• The model offers a valuable solution for reducing computational complexity, paving the way for practical intelligent tea-picking applications.