| 基于机器视觉和BP神经网络的超级杂交稻穴播量检测 |
| |
| 引用本文： | 谭穗妍,马旭,吴露露,李泽华,梁仲维.基于机器视觉和BP神经网络的超级杂交稻穴播量检测[J].农业工程学报,2014,30(21):201-208. |
| |
| 作者姓名： | 谭穗妍 马旭 吴露露 李泽华 梁仲维 |
| |
| 作者单位： | 1. 华南农业大学工程学院，广州 510642; 华南农业大学理学院，广州 510642
2. 华南农业大学工程学院,广州,510642 |
| |
| 基金项目： | 国家高技术计划（863）课题资助（2012AA10A501）；国家自然科学基金资助项目（51275209）；高等学校博士学科点专项科研基金（20114404110004）；现代农业产业技术体系建设专项资金资助（CARS-01-33） |
| |
| 摘 要： | 为了保证秧盘上每穴超级稻种子数量一致,实现精密播种作业,需对播种性能进行准确检测,但超级杂交稻播种到秧盘中,多粒种子存在粘连、重叠、交叉等情况,传统的面积、分割算法对上述情况播种量检测精度低,因此需提高上述情况种子播种量检测精度。考虑到种子连通区域的形状特征反映种子数量,该文提出一种基于机器视觉和BP神经网络超级杂交稻穴播量检测技术。针对超级稻颜色特征,采用RGB图像中红色R和蓝色B分量组成的2×R-B分量图和固定阈值法获取二值图像;投影法定位秧盘目标检测区域和秧穴;提取连通区域10个形状特征参数,包括面积、周长、形状因子、7个不变矩,建立BP神经网络超级稻数量检测模型,检测连通区域为碎米/杂质、1、2、3、4和5粒以上6种情况;试验结果表明,6种情况的检测正确率分别为96.6%、99.8%、97.2%、92.5%、86.0%、94.3%,平均正确率为94.4%,每幅图像平均处理时间0.823s,满足精密育秧播种流水线在线检测要求;研究结果为实现精密恒量播种作业提供参考。
|
| 关 键 词： | 神经网络 图像处理 种子 超级杂交稻 机器视觉 连通区域 形状特征 穴播量 |
| 收稿时间： | 2014/8/11 0:00:00 |
| 修稿时间： | 2014/10/30 0:00:00 |
Estimation on hole seeding quantity of super hybrid rice based on machine vision and BP neural network |
| |
| Tan Suiyan,Ma Xu,Wu Lulu,Li Zehua and Liang Zhongwei.Estimation on hole seeding quantity of super hybrid rice based on machine vision and BP neural network[J].Transactions of the Chinese Society of Agricultural Engineering,2014,30(21):201-208. |
| |
| Authors: | Tan Suiyan Ma Xu Wu Lulu Li Zehua and Liang Zhongwei |
| |
| Institution: | 1. College of Engineering, South China Agricultural University, Guangzhou 510642, China2. College of Science, South China Agricultural University, Guangzhou 510642, China;1. College of Engineering, South China Agricultural University, Guangzhou 510642, China;1. College of Engineering, South China Agricultural University, Guangzhou 510642, China;1. College of Engineering, South China Agricultural University, Guangzhou 510642, China2. College of Science, South China Agricultural University, Guangzhou 510642, China;1. College of Engineering, South China Agricultural University, Guangzhou 510642, China |
| |
| Abstract: | Abstract: Super hybrid rice that is widely cultivated in China is mainly planted by transplanting techniques. In super hybrid rice transplanting, a seedling nursery is a preliminary but critical one. Complete rice seedling nursing operations include seed selection, seed germination, sowing of seeds onto seeded trays, and seedling hardening. The sowing process is performed on an automated seeder sowing test line. Matured seedlings are transplanted to the field by transplanters, which are specially designed for transplanting rice seedlings.In practice, some problems arise during the sowing process. Because super hybrid rice has strong tiller ability, it requires a precise and low seeding rate, which needs to ensure 2-3 grains in each cell in tray plugs. Furthermore, rice seed traits, such as length, shape, moisture, and weight change during the sowing process, which greatly affects the performance of the seeder sowing machine. As a result, seed distribution on the tray plugs is uneven, and cavity and single grain rates in tray plug cells are high, which is up to 20%. The low seeding rate will lead to a low seedling survival rate. Also, the estimation accuracy of grains that are overlapped and adhered is quite low. To solve the problem, it is necessary to add nursery cell tray sowing quantity estimation to an automated rice sowing test line. When cavities and single grains in tray cells are detected, artificial reseeding and working parameter adjustment to the seeder sowing machine are needed.After rice seeds are sowed onto the nursery trays, seed connected regions extracted from the acquired image may occur as the following situations: impurity, single grain, and grains that are overlapped and adhered. To a great extent, the shape features of each connected region can determine the grain quantity. In this paper, a method was presented to estimate the sowing quantity per cell in the tray plug, based on machine vision and a BP neural network. The method consisted of four stages: image acquisition and preprocessing, target detection area and cell plug located, seed connected region feature extraction, and BP neural network classification. First, a series of pretreatments to the image were applied, including smoothing, denoising, image binarization, and segmentation. Second, a map projection method of a binary image was performed to locate the target detection area and cell plug. Third, three shape geometric features and seven invariant moments of each seed connected regions were extracted, which were used for inputs of a BP neural network and grain quantity estimation. And last, a BP neural network classifier was employed to classify seed connected regions into impurities, one particle, two particles, three particles, four particles, or five particles and above. After BP neural network classification was done, the seedling tray sowing quantity of each plug cell can be obtained. The result showed that the overall accuracy of the BP neural network was up to 94.4% and the classification accuracy of impurity, one particle, two particles, three particles, four particles, five particles and above were 96.6%, 99.8%, 97.2%, 92.5%, 86.0%, and 94.3% respectively. Each image was processed less than 0.823 seconds. It's proved that the method of nursery cell tray sowing quantity estimation meets the requirement of an automated rice sowing test line. |
| |
| Keywords: | neural networks image processing seed super hybrid rice machine vision connected region shape geometric features sowing quantity per cell |
| 本文献已被 CNKI 万方数据 等数据库收录！ |
| 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 |
|
