YOLO-MEST: A re-parameterized multi-scale fusion model with enhanced detection head for high-accuracy tea bud detection

Abstract

Accurate recognition of tea buds is essential for automated harvesting. However, conventional image-based methods have difficulty with complex field conditions, such as varying illumination, occlusion, and cluttered backgrounds. To overcome these challenges, we introduce YOLO-MEST, a new detection model that incorporates multi-scale feature fusion. This model introduced the RepNCSELAN4 module to enhance feature extraction capabilities and the SPPELAN module to improve feature fusion. The efficient detection head, LiteShiftHead, was added to the network output to improve the accuracy of bounding boxes and classification regression. An improved loss function dDIoU based on the difference in the width-to-height ratio of bounding boxes was designed to enhance the accuracy of bounding box localization further. To form a complete detection-to-picking pipeline, a morphological algorithm-based tea bud picking point estimation algorithm was further proposed, which effectively determined the tea bud picking points based on the detection results. We conducted performance tests on tea bud recognition based on a self-built dataset of high-quality tea. Compared to the original YOLOv8 model, the YOLO-MEST model increased mAP50 by 1.7% and mAP by 4.4%, respectively. Ablation studies confirm the contribution of each component. The proposed method significantly improves detection accuracy and supports practical intelligent tea harvesting.