| 苹果品质动态无损感知及分级机器手系统 |
| |
| 引用本文： | 彭彦昆,孙晨,赵苗.苹果品质动态无损感知及分级机器手系统[J].农业工程学报,2022,38(16):293-303. |
| |
| 作者姓名： | 彭彦昆 孙晨 赵苗 |
| |
| 作者单位： | 1. 中国农业大学工学院，北京 100083；2. 国家农产品加工技术装备研发分中心，北京 100083 |
| |
| 基金项目： | 国家重点研发计划项目( 2016YFD0400905-05) |
| |
| 摘 要： | 为了实现灵活高效的苹果多品质指标检测分级，基于机器视觉技术及可见/近红外光谱技术，开发了用于苹果内外部品质无损感知及分级的机器手系统。机器手系统采用六轴机械臂搭载自行研发的末端执行器，末端执行器上装载有光学传感器与抓取结构，可以抓取流水线上的苹果并同时采集苹果的光谱进行糖度检测。使用CMOS相机采集苹果图像，训练并使用PP-YOLO深度学习目标检测模型处理采集的苹果图像，计算苹果的坐标位置实现苹果的动态定位，并获取苹果的果径大小、着色度信息实现外部品质检测。采集苹果样本光谱，结合不同的光谱预处理方式，利用偏最小二乘（Partial Least-Square，PLS）方法进行建模分析。试验结果表明，使用PP-YOLO目标检测算法处理图像和计算苹果位置，其识别速度为38帧/s，极大地提高了检测速度。使用归一化光谱比值法（Normalized Spectral Ratio，NSR）作为预处理算法的糖度建模结果较佳。采用NSR+CARS（Competitive Adaptive Reweighted Sampling，竞争性自适应重加权算法）作为机器手的动态光谱模型效果较佳，该动态光谱模型相关系数Rv为0.958 9，验证均方根误差RMSEV（Root Mean Squared Error of Validation）为0.462 7%，与静态下建立的模型相比，机器手在动态状态下采集光谱对所建立的预测模型的预测效果影响较小。对整体机器手系统进行了试验验证，机器手在工作时能够无损伤地抓取苹果，给出果径大小、着色度、糖度3个检测指标并依据指标自动划分等级，然后依据等级信息分级。随后测定了3个指标的实测值与预测值进行分析，果径大小的预测相关系数为0.977 2，均方根误差为1.631 5 mm；着色度的预测相关系数为0.967 4，均方根误差为5.973 4%；糖度的预测相关系数为0.964 3，均方根误差为0.504 8%，预测结果与真实值均具有较强的线性关系和较低的预测误差，机器手系统分级正确率为95%，完成一颗苹果的定位、抓取、检测、分级和放置的时间约为5.2 s，具有较好的工作可靠性，研究结果为苹果多品质指标的高效检测提供参考。
|
| 关 键 词： | 机器视觉 可见/近红外光谱 苹果 无损感知 分级 机器手系统 |
| 收稿时间： | 2022/4/24 0:00:00 |
| 修稿时间： | 2022/8/7 0:00:00 |
Dynamic nondestructive sensing and grading manipulator system for apple quality |
| |
| Peng Yankun,Sun Chen,Zhao Miao.Dynamic nondestructive sensing and grading manipulator system for apple quality[J].Transactions of the Chinese Society of Agricultural Engineering,2022,38(16):293-303. |
| |
| Authors: | Peng Yankun Sun Chen Zhao Miao |
| |
| Institution: | 1. College of Engineering, China Agricultural University, Beijing 100083, China; 2. National R&D Center for Agro-processing Equipment, Beijing 100083, China |
| |
| Abstract: | Abstract: Flexible and efficient detection and classification were here proposed for the multiple quality index of the apple. In this study, a manipulator system was also developed with nondestructive sensing and grading for the internal and external quality of apple using machine vision and visible and Near-Infrared (Vis/NIR) spectroscopy. A six-Degree of Freedom (DOF) mechanical arm was used to equip a self-designed end effector in the system. Specifically, the end effector was loaded with the optical sensor and grasping structure, in order to capture the Vis/NIR spectrum of the apple. A manipulator was obtained to combine the end effector with the mechanical arm. The apple was first grabbed on the assembly line, and then the spectrum of the apple was collected at the same time for sugar content detection. The spectra of apple samples were collected in the static and dynamic states. Some spectral preprocessing was implemented for the modeling and analysis using the Partial Least Squares (PLS). A CMOS camera was selected to collect the images for the dynamic positioning and external quality detection of apples. A target detection model of PP-YOLO deep learning was trained on the apple images to calculate the coordinate position of the apple for the fruit diameter and coloration. The experimental results show that the Normalized Spectral Ratio (NSR) preprocessing performed the best in the static and dynamic states. The best performance was achieved in the dynamic spectral model of the manipulator using the NSR and Coherent Anti-Stokes Raman Scattering (CARS). The correlation coefficient, Rv, was 0.958 9 in the dynamic spectral model, where the Root Mean Square Error (RMSE) was 0.462 7%. There was less influence on the prediction model. The overall manipulator system was verified in the field test. The manipulator was used to flexibly grab the apples without damage during work. Three detection indicators were also given for the fruit diameter, coloring degree, and sugar content. An automatic grading was then implemented, according to the indicators. As such, the apples were finally placed into the corresponding level box in terms of the grade information. A comparison was also made between the measured and predicted values of the three indexes. The predicted correlation coefficient of apple diameter, coloring degree, and sugar content were 0.977 2, 0.967 4, and 0.964 3, respectively, with the RMSE of 1.631 5 mm, 5.973 4%, and 0.504 8%, respectively. There was a strong linear relationship between the prediction and actual value, indicating a lower prediction error than before. The maximum classification accuracy was up to 95% in the manipulator system. The grading system of the mechanical arm was taken about 5.2 s to realize the positioning, grasping, detection, classification, and placement of an apple, indicating better reliability. |
| |
| Keywords: | machine vision Vis/NIR spectroscopy apple nondestructive sensing grading manipulator system |
|
| 点击此处可从《农业工程学报》浏览原始摘要信息 |
| 点击此处可从《农业工程学报》下载免费的PDF全文 |
|
