Lightweight Tea Shoots Detection Algorithm Based on Deep Learning

In machine tea picking, the complex tea garden environment and frequent occlusions require an algorithm that can accurately identify tea shoots while being deployable on resource-limited mobile devices. To address this, a lightweight YOLO-TD (You Only Look Once-Tea Detect) tea shoots detection algorithm based on improved YOLOv8 is proposed. The algorithm uses lightweight C2f-faster modules for feature extraction, reducing model complexity while enhancing central feature extraction capabilities. Inner-SIoU is employed as the loss function to optimize training and improve sensitivity to small tea shoots. Additionally, the G-Head structure with a dual-branch feature extraction design is used. Experimental results show that the proposed algorithm achieves a mean average precision (mAP) of 82.20%, reduces the Giga Floating-point Operations Per Second (GFLOPS) by 45.24%, accurately identifies key points of tea shoots, and effectively detects occluded tea shoots. This improved algorithm offers a robust solution for tea shoot detection in machine tea picking.