【作者】 张毅；

【导师】 徐胜勇；

【作者基本信息】 华中农业大学 ， 农业机械化工程， 2023， 硕士

【摘要】 在工厂化育苗中,在种苗移栽阶段做出正确的分苗补苗决策,对于提高种苗质量、提升经济效益具有重要意义。目前,我国分苗补苗作业的决策仍以人工经验为主,效率低且主观性强,急需智能化的检测技术。本文提出一种基于表型检测与机器学习技术的西瓜幼苗表型检测、表型预测及弱苗判别方法。主要研究内容和结论如下:1.提出了一套Azure Kinect相机的标定方法。西瓜幼苗外形矮小,对图像采集设备要求较高,为提高使用Azure Kinect相机的幼苗表型测量精度,对彩色和深度相机进行标定。使用张正友标定法对彩色相机进行标定并对畸变进行矫正;使用深度图像像素矫正法建立深度值与真实值之间的映射关系,实现对深度相机的标定与矫正。实验结果表明彩色图像畸变得到了很好的矫正,相机在1m距离处深度信息的标准差矫正前为0.0535,矫正后为0.0056,标准差明显减小,数据更加缓和,深度信息畸变现象在一定程度上被矫正,为后续测量微小的西瓜钵苗提供了硬件保障。2.采集了早佳84-24品种、子叶展平期后连续8天的西瓜幼苗图像、人工测量的表型、壮苗指数等实验数据。使用Azure Kinect每天分三次对整盘西瓜幼苗俯视图进行数据采集,拍摄时间选择9:00、14:00以及19:00,共拍摄8天,并在数据采集的第8天根据壮苗指数方程对每株幼苗的形态指标进行人工判定。大量的实验数据为后续的算法开发和测试提供了数据保障。3.提出了西瓜幼苗的株高、叶面积无损检测和预测算法。对采集的数据进行预处理、图像分割、点云处理方法等实现对每株幼苗的定位并获取每株幼苗的株高及叶面积数据,穴盘苗穴孔识别在第1天、第2天、第3天的识别准确率分别为93.75%、87.5%、87.5%。表型检测实验中,株高与叶面积总体测量的R~2分别为0.901和0.922。对前三天的株高和叶面积使用LSTM循环神经网络进行预测推演,株高特征第4天、第5天、第6天、第7天、第8天的预测R~2分别为0.952、0.942、0.932、0.923、0.866;叶面积特征第4天、第5天、第6天、第7天、第8天的预测R~2分别为0.926、0.919、0.901、0.882、0.862。通过分析选取预测第6天数据,既可保证预测精度,又可满足育苗工厂所期望的早期预测需求。4.基于表型预测数据和人工计算的壮苗指数,使用多种分类器建立了弱苗识别模型。分别使用随机森林、SVM、XGBoost等6种机器学习分类方法对最后1天的数据进行正常苗(壮苗)和异常苗(弱苗)的二分类判别,实验结果表明LSTM和随机森林的联合预测分类模型的分类准确率最高,达到89.9%,具有良好的分类精度。5.基于PyQt5开发了一款人机软件并将西瓜幼苗表型测量算法和弱苗早期识别算法集成。该软件具有西瓜幼苗穴盘穴孔识别、单株幼苗定位、幼苗表型检测以及壮弱苗判别等功能。用户只需输入拍摄到的整盘西瓜幼苗图像,软件系统即可给出算法相对应结果与决策。人机界面友好简洁,操作方便。本文通过对西瓜幼苗表型的无损测量与预测及壮苗分类,综合计算机视觉方法和多种机器学习方法,建立了高精度、低成本的西瓜幼苗弱苗的早期判别模型。该研究可为育苗工厂实现智能无人化育苗和移栽机器人无人作业提供一定的视觉技术支持和决策依据,具有良好的推广应用价值。更多还原

【Abstract】 In factory nursery,it is important to make the correct decision of seedling division and replenishment at the seedling transplanting stage to improve the quality of seedlings and enhance the economic benefits.At present,the decision of seedling seeding replenishment operation in China is still based on manual experience,which is inefficient and subjective,and an intelligent detection technology is urgently needed.In this paper,we propose a method for phenotype detection,phenotype prediction and weak seedling discrimination of watermelon seedlings based on computer vision and machine learning technology.The main research contents and conclusions are as follows:1.A set of calibration methods for the Azure Kinect camera is proposed.To improve the accuracy of seedling phenotype measurement using the Azure Kinect camera,color and depth cameras were calibrated for watermelon seedlings with short shapes and high requirements for image acquisition equipment.The color camera was calibrated and the aberrations were corrected using the Zhang Zhengyou calibration method;the depth camera was calibrated and corrected using the depth image pixel correction method to establish a mapping relationship between the depth value and the true value.The experimental results show that the color image aberration is well corrected,the standard deviation of the depth information of the camera at 1m distance is 0.0535 before correction and 0.0056 after correction,the standard deviation is obviously reduced,the data is more moderate,the depth information aberration phenomenon is corrected to a certain extent,which provides hardware guarantee for the subsequent measurement of tiny watermelon potted seedlings.2.Experimental data such as images of watermelon seedlings,manually measured phenotypes,and strong seedling indices were collected for 8 consecutive days after the cotyledon spreading and flattening stage of the early Jia 84-24 variety.The Azure Kinect was used to collect data from the top view of the entire tray of watermelon seedlings three times a day at 9:00,14:00 and 19:00 for eight days,and the morphological index of each seedling was manually determined based on the robustness index equation on the eighth day of data collection.The large amount of experimental data provided data for the subsequent algorithm development and testing.Experimental data such as images of watermelon seedlings of early Jia 84-24 variety,cotyledon spreading stage after 8 consecutive days,manually measured phenotypes,and strong seedling index were collected.The Azure Kinect was used to capture the top view of the entire tray of watermelon seedlings three times a day,at 9:00,14:00 and 19:00,for eight days,and the morphological index of each seedling was manually determined based on the robustness index equation on the eighth day of data collection.The large amount of experimental data provided data for the subsequent algorithm development and testing.3.A nondestructive detection and prediction algorithm for plant height and leaf area of watermelon seedlings is proposed.The pre-processing,image segmentation and point cloud processing methods were used to locate each seedling and obtain plant height and leaf area data for each seedling.93.75%,87.5% and 87.5% of the recognition accuracy were obtained for hole hole identification of cavity tray seedlings on day 1,day 2 and day 3,respectively.In the phenotype detection experiment,the overall R2 of plant height and leaf area measurements were 0.901 and 0.922,respectively,and the predicted R2 of plant height and leaf area for the first three days using LSTM recurrent neural network were 0.952,0.942,0.932,0.923,and 0.866 for plant height,t+1,t+2,t+3,t+4,and t+5,respectively.The R2 for leaf area at t+1,t+2,t+3,t+4,and t+5 were0.926,0.919,0.901,0.882,and 0.862,respectively.t+3 was chosen as the step model to ensure the prediction accuracy and to meet the desired early prediction needs of the nursery plant.4.Based on the phenotypic prediction data and the manually calculated strong seedling index,a weak seedling identification model was established using multiple classifiers.Six machine learning classification methods,including random forest,SVM,and XGBoost,were used to discriminate normal seedlings(strong seedlings)and abnormal seedlings(weak seedlings)dichotomously on the last day of data,and the experimental results showed that the joint predictive classification model of LSTM and random forest had the highest classification accuracy of 89.9%,with good classification accuracy.5.A human-machine software based on PyQt5 was developed and integrated with watermelon seedling phenotype measurement algorithm and weak seedling early identification algorithm.The software has the functions of watermelon seedling cavity identification,single seedling location,seedling phenotype detection,and strong and weak seedling identification.The user only needs to input the image of the whole tray of watermelon seedlings,and the software system can give the algorithm corresponding results and decisions.The human-machine interface is friendly,simple and easy to operate.This paper establishes a high-precision and low-cost model for early discrimination of weak watermelon seedlings by non-destructive measurement and prediction of watermelon seedling phenotype and classification of strong seedlings,and integrates computer vision methods and various machine learning methods.The research can provide some vision technical support and decision basis for the realization of intelligent unmanned seedling plant and unmanned operation of transplanting robot in seedling plant,and has good promotion application value.更多还原