Ru Han ¹ Ye Zheng ¹ Renjie Tian ¹ Lei Shu ^1* Xiaoyuan Jing ² Fan Yang ³

• ¹ Nanjing Agricultural University, Nanjing, China
• ² Guangdong University of Petrochemical Technology, Maoming, Guangdong, China
• ³ Jiangsu Normal University, Xuzhou, Jiangsu Province, China

As a global leader in tea production, China holds tea not just as an important cash crop, but also as a product leading in market sales worldwide. Tea plantation harvesting stands at the heart in the comprehensive process of tea production, with tea picking as 1 Han et al. a critical component that is increasingly moving toward intelligence and mechanization (Kisantal et al., 2019). By leveraging smart recognition technologies, we are capable of conducting precise monitoring and management of picking operations in large tea plantation. Specifically:• To align with the standardized requirements for tea operation procedures and harvesting management, tea plantation can utilize smart recognition technologies to monitor in real time and rectify inappropriate picking behaviors, such as excessive picking or irrational leaf handling, ensuring the stabilization and enhancement of tea quality.• In tourist-accessible tea plantation, administrators can employ intelligent surveillance systems to ensure that visitors' picking activities adhere to regulations, maintaining order within the gardens. For those rare ancient tree tea plantation, smart recognition technologies serve as a potent instrument to prohibit picking and protect precious plant specimens. In production-oriented tea plantation, similar monitoring of farmers' picking behaviors can prevent irregular practices, safeguarding production efficiency and tea quality. (1) This paper selected videos about tea picking on the Internet, and crawled most of the tea picking videos on searchable platforms such as Baidu,Good-looking video, Watermelon video, and Bilibili, and sliced the videos. The video scenes include various terrains such as mountains, hills, and plains, and also include various weather conditions: 1) sunny; 2) overcast; 3) cloudy; 4) foggy; 5) rainy. Han et al. The image video resolutions include 480p, 720p, and 1080p formats, and the video slice sizes include 1920×1280, 854×480, 852×480.(2) The tea picking behavior recognition dataset constructed in this work has a total of seven categories of labels: 1) picking; 2) picking (machine); 3) walking; 4) standing; 5) talking; 6) storage tools; 7) picking tools. Considering that some original images do not contain any of the aforementioned categories during the image slicing process, such images may affect the final judgment. Therefore, while ensuring the continuity of the video as much as possible, we deleted some error-prone image frames and sorted the annotated images to serve as the raw data for the dataset. The Figure 1 The abnormal images that need to be processed mainly include the following situations:• When the tea bushes block the picking behavior in large areas, it is difficult to recognize the picking behavior in the image. Han et al. • When the shooting distance is too far, the targets in the image are too small to distinguish and recognize.• Due to limited or blocked shooting angles and scenes, the image data obtained after slicing only includes partial features of the labels, resulting in low recognition accuracy.During the processing of the image dataset, manual or automatic deletion of the above abnormal images is required. At the same time, videos with the same content but different titles on the internet or videos that are contained in other videos are also filtered and deleted. Based on the original data, we performed data augmentation operations to expand the dataset and provide more training data.The basic data augmentation methods include rotation, flipping, enhancement, and cropping.The Table 1 details and lists the data augmentation methods and their corresponding quantities. The Figure 3 Han et al. 2) Fewer behavioral features: Most of the small object labels are related to behavior recognition (such as picking tools, storing tools, etc.), which results in less information related to the features of the small objects themselves. Behavior recognition relies more on the overall posture of the object and the surrounding environment rather than subtle details. The annotation work is carried out using the labelme software. As shown in Table 3, We evaluated the performance of three popular object detection algorithms (Faster R-CNN, SSD, and YOLOv5s) on the tea garden picking dataset. By comparing their precision (P), recall (R), and mean average precision at 50% IoU (mAP50), we can come to the following conclusions:1) YOLOv5s performed exceptionally well across all three metrics, achieving the highest precision (84.5%), recall (78.8%), and mAP50 (82.3%). This indicates that YOLOv5s not only accurately identifies targets in tea garden picking scenarios but also covers a larger number of true targets while maintaining high overall detection performance.2) SSD exhibited the best precision (81.90%) but fell short in recall (66.80%), resulting in an mAP50 of 72.95%. This suggests that SSD might miss some targets in certain situations, despite its high accuracy in identifying targets.3) Faster R-CNN excelled in recall (81.72%) but had a relatively lower precision (67.56%), leading to more false positives. Its mAP50 of 79.67% indicates challenges in balancing precision and recall.In summary, YOLOv5s balanced precision and recall best on the tea garden picking dataset, suitable for high-performance applications. Frontiers