✦
INTRODUCTION
✦
Accurate and real-time wheat ear counting in the field is of great significance for wheat yield prediction, genetic breeding and optimal planting management. The large resolution of UAV images, the large number of wheat ears in each frame image, and the small pixels of wheat ears in the image bring difficulties and challenges to the research of wheat ear counting. The drone video stream wheat ear image obtains wheat ear information from different angles, which can more accurately observe each wheat ear. Therefore, based on the above objectives, we constructed the wheat ear detection model YOLOv7xSPD based on YOLOv7x, and constructed the video stream wheat ear counting model YOLOv7xSPD Counter based on YOLOv7xSPD and Kalman filter tracking algorithm.
✦
Text
✦
To evaluate the accuracy of YOLOv7x SPD in ear counting under video streaming, we compared the accuracy (P), recall (R), F1 score (F1) and average precision (AP) of six common models, and the average accuracy of YOLOvSPD reached the highest 94.99% as shown in Table 1 and Figure 1.
表1 描述Faster RCNN、SSD、YOLOv5s、YOLOv7、YOLOv7x和YOLOv7xSPD的Precision, Recall, F1 Score, 和 AP评估结果
图1 a. Faster RCNN、SSD、YOLOv5s、YOLOv7、YOLOv7x与YOLOv7xSPD的PR曲线,b. 局部放大图,展示了图a中虚线框部分的放大效果,能清晰的表示YOLOv7与YOLOv7xSPD的PR曲线之间的差异
对比了YOLOv7x和YOLOv7xSPD的检测测效果,如图2所示,YOLOv7xSPD能够弥补YOLOv7x在麦穗尺寸较小时出现的部分漏检。
图2 YOLOv7xSPD与YOLOv7x检测效果对比。 a. 原图,b. YOLOv7x检测效果,c. YOLOv7xSPD检测效果。 b中放大部分的红色框表示YOLOv7x未检测到示YOLOv7x未检测到的麦穗,而在c中均能被精准的检测出来
The coefficient of determination (R2) and root mean square error (RMSE) were used to evaluate the counting performance of the six models, and the regression analysis of the counting results was performed, as shown in Figure 3. The YOLOv7xSPD count reached an R2 of 0.99 and an RMSE of only 3.39.
图3 真实数量与Faster RCNN、SSD、YOLOv5s、YOLOv7、YOLOv7x和YOLOv7xSPD的计数结果的相关性
在对YOLOv5s、YOLOv7、YOLOv7x和YOLOv7SPD计数结果的统计中,YOLOv7SPD的绝对误差更小,如图4所示。
图4 YOLOv5s、YOLOv7、YOLOv7x、YOLOv7xSPD的计数结果与真实值的差距数的直方图(a)和密度(b)
The YOLOv7xSPD Counter also achieved a count of 0.99 in the video stream, with an RMSE of only 10.50, and 5.5 frames at a resolution of 3840×2160, as shown in Figure 5.
Fig.5 YOLOv7xSPD Counter was used to perform regression analysis, RMSE and R2 calculations on the counting results of 20 video test sets
Source:
✦
Xu, XM., Zhou, L., Yu, HL., Sun, GY., Fei, SP., Zhu, JY*., Ma, YT*. Winter wheat ear counting based on improved YOLOv7x and Kalman filter tracking algorithm with video streaming. Frontiers in Plant Science. 2024; 15. https://www.frontiersin.org/journals/plant-science/articles/10.3389/fpls.2024.1346182/full
About the Author
✦
The first author of the paper is Associate Professor Xu Xingmei from the College of Information Technology of Jilin Agricultural University, and the corresponding author of the paper is Professor Ma Yuntao from the College of Land Science and Technology of China Agricultural University, and the collaborators include Professor Yu Helong and Zhou Lei from the College of Information Technology of Jilin Agricultural University. The main research directions of the digital agriculture innovation team of China Agricultural University are system simulation and digital twin of plant function-structure-environment interaction, data mining and application of plant growth information based on machine vision, rapid investigation of large-scale breeding traits of unmanned aerial vehicles, creation of breeding robots and agricultural sensors, multi-source sensor fusion and digital agriculture applications, artificial intelligence and smart agriculture, etc. Long-term enrollment of a number of master's, doctoral students and cooperative postdoctoral fellows, interested parties please contact: [email protected].
Source: CAU Digital Agriculture Team
Wheat multi-omics website: http://wheatomics.sdau.edu.cn submission, cooperation, etc. E-mail: [email protected] WeChat group: Click on the exchange group of the wheat research alliance to join the group