基于虚拟数据和旋转目标检测分析的大豆豆荚表型参数测量方法  

Measurement method for soybean pod phenotypic parameters based on virtual data and rotated object detection analysis

作　　者：吴康磊 金秀[1,2] 饶元 李佳佳[3] 王晓波 王坦[1,2] 江朝晖[1,2] WU Kanglei;JIN Xiu;RAO Yuan;LI Jiajia;WANG Xiaobo;WANG Tan;JIANG Zhaohui(College of Information and Artificial Intelligence,Anhui Agricultural University,Hefei 230036,China;Key Laboratory of Agricultural Sensors,Ministry of Agriculture and Rural Affairs,Hefei 230036,China;College of Agronomy,Anhui Agricultural University,Hefei 230036,China)

机构地区：[1]安徽农业大学信息与人工智能学院,安徽合肥230036 [2]农业农村部农业传感器重点实验室,安徽合肥230036 [3]安徽农业大学农学院,安徽合肥230036

出　　处：《江苏农业学报》2024年第7期1245-1259,共15页Jiangsu Journal of Agricultural Sciences

基　　金：国家自然科学基金项目(32371993);安徽省重点研究与开发计划项目(202204c06020026、2023n06020057);安徽省高校自然科学研究重大项目(2022AH040125、2023AH040135)。

摘　　要：为解决传统大豆考种过程中人工测量大豆豆荚表型参数耗时费力的问题以及现有的自动化测量方式存在的人工数据标注需求量大、环境适应能力弱、计算代价高等问题,本研究提出一种基于虚拟数据集生成和旋转目标检测分析的豆荚关键表型参数自动化测量方法,重点关注荚长和荚宽的测量。该方法基于YOLOv7-tiny提出一种改进的豆荚检测模型(CSL-YOLOv7-tiny),通过引入环形平滑标签使模型获得对旋转目标的检测能力,提升对无序摆放的狭长豆荚目标检测的质量。为避免人工标注训练数据,采用虚拟图像生成方法得到含标注信息的虚拟豆荚数据集和虚拟硬币与豆荚混合数据集。利用迁移学习策略,将模型从虚拟豆荚数据集迁移至虚拟硬币与豆荚混合数据集,积累模型对豆荚特征的提取能力。设计一种基于K-均值聚类的后处理方法,对检测到的旋转边界框进行分析,得到荚长和荚宽,以减少拍摄环境差异带来的测量误差。试验结果表明,在无任何训练数据标注的条件下,使用虚拟图像训练的CSL-YOLOv7-tiny对硬币和豆荚目标检测的最优mAP_(0.50)和mAP_(0.50∶0.95)分别达到了99.3%和78.0%,其模型大小和推理时间分别仅为12.92 MB和12.5 ms,荚长和荚宽测量的决定系数(R^(2))分别达到了0.94和0.86,与实际测量均值分别仅相差0.42 mm和0.02 mm。此外,通过对本研究提出的方法进行对比分析,验证了其在模型训练、轻量化部署以及不同考种环境适应能力上的优势。研究结果可为大豆豆荚表型参数的自动化、智能化测量系统的研发提供参考,为加速优质高产大豆的选育进程提供支撑。To solve the problems such as time-consuming and labor-intensive of manual measurement for soybean pod phenotypic parameters in traditional soybean seed evaluation processes,as well as the large quantity demand for manual data annotation,weak environmental adaptation and high computational costs in existing automated measurement methods,an automated measurement method for pod key phenotypic parameters which was mainly focused on pod length and width measuring was proposed in this study,based on virtual dataset generation and rotated object detection analysis.An improved pod detection model(CSL-YOLOv7-tiny)was proposed by the method based on YOLOv7-tiny.The Circular Smooth Label was introduced to enable the model to obtain the capability for rotated object detection,and to improve the quality of detecting elongated pod targets in a disorganized arrangement.To avoid manual annotation of training data,virtual image generation method was used to get virtual pod dataset as well as virtual coin and pod mixture dataset containing annotation information.Transfer learning strategy was employed to transfer the model from the virtual pod dataset to the virtual coin and pod mixture dataset,which accumulated the model’s ability in pod features extracting.A post-processing method based on K-means clustering was designed to analyze the detected rotated bounding boxes,and obtained pod length and width,which reduced measurement errors caused by shooting environmental differences.Experimental results showed that under the condition of no training data annotation,CSL-YOLOv7-tiny trained by virtual images obtained the optimal mAP_(0.50)and mAP_(0.50∶0.95)for coin and pod targets detection,which were 99.3%and 78.0%,respectively.The model size and inference time were only 12.92 MB and 12.5 ms respectively,and the determination coefficients(R^(2))for pod length and width measurement reached 0.94 and 0.86 respectively,with only 0.42 mm and 0.02 mm differences compared with actual measurements.Furthermore,by comparative analysis of t

关 键 词：大豆考种 豆荚表型 虚拟数据 旋转目标检测 YOLOv7-tiny 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]