TS-YOLO: An All-Day and Lightweight Tea Canopy Shoots Detection Model

Abstract

Accurate and rapid detection of tea shoots within the tea canopy is essential for achieving the automatic picking of famous tea. The current detection models suffer from two main issues: low inference speed and difficulty in deployment on movable platforms, which constrain the development of intelligent tea picking equipment. Furthermore, the detection of tea canopy shoots is currently limited to natural daylight conditions, with no reported studies on detecting tea shoots under artificial light during the nighttime. Developing an all-day tea picking platform would significantly improve the efficiency of tea picking. In view of these problems, the research objective was to propose an all-day lightweight detection model for tea canopy shoots (TS-YOLO) based on YOLOv4. Firstly, image datasets of tea canopy shoots sample were collected under low light (6:30–7:30 and 18:30–19:30), medium light (8:00–9:00 and 17:00–18:00), high light (11:00–15:00), and artificial light at night. Then, the feature extraction network of YOLOv4 and the standard convolution of the entire network were replaced with the lightweight neural network MobilenetV3 and the depth-wise separable convolution. Finally, to compensate for the lack of feature extraction ability in the lightweight neural network, a deformable convolutional layer and coordinate attention modules were added to the network. The results showed that the improved model size was 11.78 M, 18.30% of that of YOLOv4, and the detection speed was improved by 11.68 FPS. The detection accuracy, recall, and AP of tea canopy shoots under different light conditions were 85.35%, 78.42%, and 82.12%, respectively, which were 1.08%, 12.52%, and 8.20% higher than MobileNetV3-YOLOv4, respectively. The developed lightweight model could effectively and rapidly detect tea canopy shoots under all-day light conditions, which provides the potential to develop an all-day intelligent tea picking platform.

Detection and Localization of Tea Bud Based on Improved YOLOv5s and 3D Point Cloud Processing

“…where m represents the centroid coordinate of the 3D point cloud, n represents the number of 3D point clouds, and P i represents the 3D coordinate of the ith point. Subsequently, the covariance matrix C is derived from the centroid coordinate, as presented in Equation (8).…”

“…Moreover, it reduced the average computational complexity by 89.11% and the number of parameters by 82.36%. Zhang et al [8] proposed using MobileNetV3 as the backbone network of YOLOv4, replacing the original convolution with a depth-separable convolution and introducing a deformable convolutional layer and a coordinate attention module, and the experimental results show that under different lighting conditions, the detection accuracy, recall, and AP are 85.35%, 78.42%, and 82.12%, respectively. Zhang et al [9] achieve the goal of reducing the model size by removing the focus layer and replacing the original feature extraction network of YOLOv5 with the ShuffleNetv2 algorithm, followed by channel pruning at the head of the neck layer, and the experimental results show that the detection speed can be up to 8.6 frames/second.…”

LSR-YOLO: A High-Precision, Lightweight Model for Sheep Face Recognition on the Mobile End

“…By embedding location information into channel attention, the CA attention mechanism enables the model to obtain information about a larger area. The CA attention mechanism contains both channel and spatial attention modules, outperforming SE with only the channel mechanism [30,31]. The structure diagram of the CA attention mechanism is shown in Figure 5.…”

Influence of tea leaf morphology on water interception for droplet impact

“…Tea ( Camellia sine n sis [L.] O. Kuntze) is one of the high‐value cash crops in hilly areas of China, and tea products have become the most consumed beverage worldwide (Zhang et al, 2023). The global climate is undergoing drastic changes and extreme weather is occurring frequently (Lesk et al, 2016).…”