基于改进YOLOv7的设施番茄苗期株高检测方法研究

An improved YOLOv7-based height detection method for seedling tomato in greenhouse

关键词:  番茄  植物表型  株高  目标检测  YOLOv7  CARAFE

Key words:  tomato  plant phenotype  plant height  object detection  YOLOv7  CARAFE

摘要:

【目的】 株高是判断番茄长势的重要指标之一，对番茄生长发育、栽培管理等研究具有重要意义。文章以设施番茄为研究对象，开展基于机器视觉和深度学习的设施番茄苗期株高检测方法研究，为实现在复杂背景下的设施番茄株高快速、无损、实时检测提供重要支持。【方法】 采用自主研发的日光温室作物表型数据采集与分析云平台，定期采集不同品种番茄苗期植株侧视图；针对平台获取到的可见光图像，进行番茄植株与参照物目标框标注的工作；在YOLOv7（You Only Look Once version7）目标检测模型基础上，改进模型上采样方法，使用CARAFE（Content-Aware ReAssembly of Features）代替原近邻插值上采样方法，获取更加丰富的特征信息；基于改进后的模型实现番茄植株和参照物等多类别目标检测并计算输出株高信息。【结果】 经过实验对比，所提出的株高检测方法获取的株高与实测值的均方根误差（RMSE）为2.06 cm，决定系数（R²）为0.87；相比基于YOLOv4模型、YOLOv5模型和YOLOv7模型计算的株高，RMSE分别减少了1.4 cm、0.85 cm和0.85 cm，R²分别提高了0.25、0.17和0.14。【结论】 该文提出的设施番茄苗期株高检测方法，可为设施环境下苗期番茄株高实时、快速、无损获取提供重要技术支撑。该研究对加强设施作物监测系统、提高农业生产力和推进温室作物管理领域的研究具有重要意义。

Abstract:

Purpose Plant height is a crucial indicator for evaluating the growth status of tomato plants. Accurate and non-destructive monitoring of plant height is of great importance for optimizing tomato cultivation management and cultivar breeding. In this study，we developed a machine vision and deep learning-based method for detecting the height of tomato plants at the seedling stage，aiming to provide technical support for rapid，real-time and non-destructive detection of tomato seedling plant height under complex backgrounds.Method This study used a custom-designed plant phenotyping platform to collect and analyze plant phenotyping information in greenhouses. This platform facilitated the regular collection of side view images of tomato plants at the seedling stage. The annotation tool was used to mark the target image of tomato plants and reference objects. A target detection model based on the YOLOv7（You Only Look Once version7）algorithm was developed and enhanced by implementing a Content-Aware ReAssembly of Features（CARAFE）up-sampling method，resulting in more detailed feature information compared to the original nearest neighbor interpolation up-sampling method. This improved model was used to detect multi-category targets，such as tomato plants and reference objects，and to calculate and output of plant height information.Result After theoretical analyses and experimental comparisons，tests were conducted on several types of tomato seedling datasets to obtain experimental results. The proposed plant height detection method achieved a RMSE（Root Mean Square Error）of 2.06 cm and a coefficient of determination（R²）of 0.87 when compared with the measured plant height. In comparison to the plant heights calculated based on YOLOv4，YOLOv5 and YOLOv7 models，the RMSE reduced by 1.4 cm，0.85 cm and 0.85 cm respectively，while the R² increased by 0.25，0.17 and 0.14 respectively.Conclusion The method proposed in this study can provide important technical support for the real-time，rapid and non-destructive detection of tomato plant height at the seedling stage in controlled environments. This study holds significant implications for improving the greenhouse plant monitoring system，strengthening agricultural productivity and advancing the research in the field of greenhouse management.