【作者】 谭穗妍；

【导师】 马旭；

【作者基本信息】 华南农业大学 ， 农业机械化， 2016， 博士

【摘要】 超级杂交稻与常规杂交稻和杂交稻相比强调少本稀植,依靠分蘖能力增加有效穗数,提高产量。所以它的播种技术需要降低播量精密播种,保证2~3粒/穴,通过提高播种均匀性,减少种子用量,节本增效,提高产量。但水稻育秧采用催芽后播种,催芽的种子在播种过程中,种子物理的特性将发生变化,会影响播种性能,亟需开展超级杂交稻秧盘育秧精密播种性能检测方法及关键技术研究,以提高播种精度,保证播种育秧质量。本文创新性地提出一种基于机器视觉和嵌入式机器视觉的秧盘育秧精密播种性能检测技术,在线动态检测连续输送的播种秧盘,对采集的图像序列采用图像拼接技术自动地拼接整盘播种秧盘图像,保留每秧穴播种图像细节信息,实现秧盘每秧穴播种量的精密检测。在秧盘超级杂交稻种子播种量检测算法方面,通过多特征提取技术和特征优选技术,提出基于模式识别超级杂交稻连通区域的播种量分类识别模型,提高多粒种子颗粒数的检测精度,特别是重叠、交叉和粘连的情况,同时研究改进的分水岭算法和SUSAN算子对超级杂交稻连通区域进行分割与计数技术。本文的主要研究内容和结果如下:(1)进行了播种秧盘图像采集系统设计及图像预处理研究。根据实际超级杂交稻播种应用的情况,针对播种秧盘图像监测场景、预期装置中相机高度及拍摄视野、稻种几何特征等,在现有的超级杂交稻育秧流水线上设计了秧盘育秧图像的采集系统;试验选择了3种具有一定物理特征区别的超级杂交稻作为本课题的试验样本;针对播种秧盘中超级杂交稻目标小及其与土壤、秧盘等背景颜色的差异,通过颜色模型分析和试验研究,提取了基于RGB颜色模型中的G分量图的稻种图像分割技术;利用最大类间方差自适应阈值的图像分割算法提取稻种的连通区域;通过研究分析比较现有的边缘提取算法,利用Canny算子提取稻种边缘,确定了实现秧盘图像的预处理流程。(2)基于BP神经网络的超级杂交稻连通区域播种量识别算法研究,试验结果表明通过特征选择,三个不同品种的超级杂交稻建立各自的模式分类器获得最高的平均检测精度,分别为91.6%、93.5%和93.1%。根据超级杂交稻播种秧盘中不同颗粒数的连通区域形态各异,从形态特征方面出发,研究了稻种连通区域的形状几何特征和不变矩特征两种特征的提取方法,在几何形状特征提取方面,采用了面积、周长、形状因子、偏心率、等效面积圆直径、延伸率,同时采用7个不变矩公式提取稻种连通区域的不变矩特征;为了降低多维特征值之间的冗余信息,提高模式分类器的识别效率,研究了稻种连通区域的特征值选择算法,根据特征选择的概念和方法,通过平均影响值MIV和BP神经网络算法,对不同特征参数组合进行评估和排序,从而完成不同超级杂交稻图像的分类特征组合的选择;最后对不同的特征值组合进行构建及选取BP神经网络作为稻种连通区域播种量模式识别的方法,对连通区域为碎米/杂质、1粒、2粒、3粒、4粒及5粒以上的6种情况进行识别分类,实现超级杂交稻播种量的检测;通过试验比较不同特征值组合的平均正确率和每个超级杂交稻品种的识别正确率,说明本研究中特征提取和特征选择方法以及模式识别方法的正确性和有效性。(3)基于改进分水岭算法和SUSAN算子的超级杂交稻分割与计数技术研究,试验表明三种超级杂交稻的平均分割准确率为93.84%、92.2%和91.9%。提出一种兼顾标记提取前的预处理和对分割后的后处理的改进分水岭分割算法和SUSAN算子实现粘连杂交稻分割和计数自动化检测。为对灰度图像中区域最小值进行准确的提取,在应用分水岭分割算法前,利用小波分析对水稻灰度图像进行图像增强处理,然后通过高斯滤波对增强的灰度图像进行平滑,使得每个水稻种子内部形成一个局部极小值,最后通过分水岭算法分割稻种连通区域,使过分割区域大大减少;对于不可避免的过分割稻种区域,研究提出通过判断稻种边缘的角点与分水岭分割线端点是否重合来检测过分割区域,并通过检测与过分割线端点最邻近的两个连通区域实现过分割区域的合并;提出一种改进的SUSAN算子检测超级杂交稻连通区域轮廓的角点,包括凹点和凸点,试验结果表明:采用三种典型的杂交稻各25幅播种图像进行测试,每幅图像的稻种数量范围约150~250个,三种超级杂交稻的平均分割准确率为93.84%、92.2%、91.9%,算法适用于三种超级杂交稻的分割和计数。(4)开展了超级杂交稻穴播量精密播种性能检测技术研究,试验结果表明秧盘育秧播种合格率、空穴率、重播率及平均粒数检测精度分别为99.36%、91.77%、90.49%和96.45%。为全面精确反映整个播种秧盘的播种性能检测的信息,本文通过采集播种秧盘图像序列采用图像拼接技术拼接整盘播种秧盘,保留每秧穴播种图像细节信息;通过分析现有的图像拼接算法,提出一种结合相位相关和SURF算法的快速播种秧盘图像拼接技术。试验结果证明,本文提出的改进图像拼接算法在匹配精度、拼接时间、拼接效果方面都优于传统SURF算法,拼配精度比传统的拼接算法提高了7.14%;在拼接时间方面,改进的算法的速度是传统拼接算法的3倍;在拼接效果方面,包括R、G、B分量图融合的灰度均方误差RMSEr,RMSEg,RMSEb和配准误差RMSEreg,改进的拼接算法比传统的拼接算法误差分别降低8.4%,6.9%,6.9%和33.8%。在研究了超级杂交稻连通区域的播种量检测算法的基础上,为了实现秧盘播种性能的精密检测,提出了基于灰度投影的播种秧盘检测区域和秧穴定位算法,给出了“穴播量”检测流程,最后统计秧盘播种性能,与人工检测秧盘播种性能相比,本文提出的算法在秧盘育秧播种合格率、空穴率、重播率及平均粒数检测精度分别为99.36%、91.77%、90.49%和96.45%。(5)基于嵌入式机器视觉的秧盘育秧图像无线传输系统设计与试验,实现秧盘图像的无线传输,试验结果表明,传输系统中嵌入式采集终端能够稳定把采集的秧盘图像上传到服务器。设计了以嵌入式开发平台Tiny4412、红外传感模块、网络高清摄像头、远程服务器组建的水稻育秧流水线播种秧盘图像无线传输系统,利用Wifi网关,嵌入式系统和远程服务器按照规定的协议通过Socket通信进行数据传输,以实现秧盘图像无线传输,并存储在远程服务器进行水稻秧盘播种性能分析。试验结果表明,在搭建的Wifi网络测试环境中,传输不同数据格式的播种秧盘图像,分别为未经过数据压缩的BMP图像和经过数据压缩的JPEG图像,传输分辨率为640像素×480像素、960像素×544像素、1280像素×720像素图像时,BMP图像平均每幅图像传输时间分别为0.45s、0.72s、1.2s,传输JPEG图像平均每幅图像传输时间分别为30ms、100ms、280ms,两种图像格式的传输速度都满足育秧流水线实时作业要求;同时传输系统中嵌入式采集终端能够稳定采集播种秧盘图像,并成功地上传到服务器,网络平均丢包率为0.23%,误码率为0.23%。更多还原

【Abstract】 Super hybrid rice is widely cultivated in china.Because super hybrid rice has strong tiller ability,it requires precise and low seeding rate,which need to ensure 2~3 grains each cell in tray plugs.But during the sowing process,some problems arise.Rice seed traits,such as length,shape,moisture,weight change,which greatly affect the performance of seeder sowing machine.As a result,seed distribution on the tray plugs is uneven and sowing quantity changes now and then.To solve the problem,a seeding performance measurement system based on computer vision and embedded computer vision is put forward.Some critical technologies are planned to study by the following methods.Firstly seeding tray images sequence are acquired on line,then image mosaic algorithm is applied to form a complete nursery tray sowing image,which retain the detail information of seeding tray image.Secondly estimation of the nursery tray sowing quantity per cell is achieved and sowing performance parameters are counted.In the aspect of rice seed counting algorithm,particles quantity detection is base on seed connected regions pattern recognition,multi-feature extraction and feature selection technologies.The method improves the counting accuracy of multiple particle seed quantity,and works well when seeds are overlapped,crossed and adhered.Research on segmentation and counting for touching hybrid rice grains based on the improved watershed agrolithm was also studied.At last the development of machine vision system ensures high-speed and efficient operation of sowing performance measurement system.The theory and method of this research will open up a new technology of precision constant seeding;also the research has great theoretical value and practical application significance.The main contents and findings are as follows:(1)The image acquisition system was designed and image preprocession was studied.According to the actual situation of the application of super hybrid rice seeding,and considering the height and shooting window of camera and physical property of super hybrid rice,a nuersy plug tray image acquisition system which was placed on the sowing test line was designed.Three types of super hybrid rice with different physical property were choosed as the experimental samples of this project.As super rice was small and the color difference between rice and soil,the Otsu threshold segmentation method based on RGB color model was applied.The Otsu algorithm was applied to the G component of seeding images for getting the threshold adaptively.Through the analysis and comparison of the existing edge extraction algorithm,canny operator was used to extract the edge of the rice connected regions.To this end,the image preprocessiong process was done.(2)Research on quantity recognition algorithm of connected regions of super hybrid rice based on BP neural netweork.The experiment result showed that after feature selection of super hybrid rice,three varieties with respective establishment of pattern classifier obtained the highes average accuracy,91.6%,93.5% and 93.1% respectively.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 great extent,the shape features of each connected region can determine the grain quantity.Six shape geometric features and seven invariant moments of each seed connected regions were extracted,which were used for inputs of BP neural network and grain quantity estimation.In order to reduce the redundant information between multi-dimension features and improve the efficiency of pattern recognition,the feature selection algorithm was studied.Based on the mean impact value MIV and BP neural network algorithm,different group of feature parameters were evaluated and sorted,so as to select the best group of feature parameters.Then,a BP neural network classifier was employed to classify seed connected region into impurities,one grain,two grains,three grains,four grains or five grains and above.The result showed that the BP neural network classifier was effectiveness.(3)Research on segmentation and counting for touching hybrid rice grains based on the improved watershed agrolithm.The results showed that the average counting accuracy of three types super hybrid rice were 93.84%,92.2% and 91.9,respectively.This study proposes an improved watershed segmentation algorithm,which is based on preprocessing and post segmentation.The study can achieve the automatic segmentation and counting of touching rice.In the preprocessing stage,an improved watershed segmentation algorithm is put forward.First,wavelet transformation is used for rice gray image enhancement which better reflect the gray difference level of rice surface and edge,and then,Gauss filter is used for image smoothing.At last,the number of over-segmentation can be greatly reduced with the application of the watershed segmentation algorithm for segmenting and counting touching rice.However,there are some over-segmentation regions that inevitably exist.In order to accurately merge the over-segmentation regions,firstly SUSAN detector is applied to detect the corners around the rice seed boundary,including convex points and concave points,and then,the watershed segmentation lines are skeletonized and the endpoints of the lines are detected.Finally,whether the endpoints of the lines are coincided with the corners are detected.If the endpoints of the lines coincide with one of the corners,the segmentation lines are judged as correct lines,in contrast,if the endpoints of lines don’t coincided with one of the corners,the segmentation lines can be judged as over-segmentation lines.The two nearest connected regions of the endpoints of over-segmentation lines are over-segmented regions,and the two regions will be merged.25 images of each kinds of hybrid rice Pei Za Tai Feng,Teyou No.338 and Tai Feng You 208 were used for the tests.The result showed that the average accuracy of segmentation and counting of Pei Za Tai Feng,Teyou No.338 and Tai Feng You 208 rice grains were 93.84%、92.2% and 91.9% respectively.(4)A method was presented to estimate the sowing quantity per cell in tray plug and precision seeding performance.The test result showed that the test accuracy of qualified seeding rate,leakage rate,reply rate and the average grain number are 99.36%、91.77%、90.49% and 96.45% respectively.After sowing quantity of connected regions of super hybrid rice was studied,map projection method of binary image was performed to locate target detection area and cell plug.In order to fully and accurately obtain the imformation of the seeding performance nursey tray,a fast nursery plug tray image mosaic technique based on phase correlation and Speeded up Robust Features(SURF)detection is introduced.The test results show that the proposed method greatly outperforms the traditional SURF method in point matching accuracy,time consumption,and image mosaic accuracy.The average feature point matching accuracy of this new method is improved by approximately 7.14%.The implementation time is almost three times faster.The Root Mean Squared Error(RMSE)of image blending in the R channel RMSEr,the G channel RMSEg,and the B channel RMSEb,are decreased by approximately 8.4%,6.9%,6.9% respectively.The Root Mean Squared Error of image registration RMSEreg is decreased by 33.8%.(5)A rice nursery tray images wireless transmission system based on embedded machine vision was designed.The embedded machine vision system was composed of embedded development platform Tiny4410,Wifi gateway,network camera,infrared sensor module and remote computer.The embedded Linux operating system,camera driver,GPIO port driver and network file system configuration were installed in embedded development platform.Applications for the device were programmed with Qt development tool.The applications included image acquisition,real-time images displayed on screen and friendly interactive interface.Jpeglib static library was used to compress the images.Through the Wifi network,embedded system and remote server achieve socket communication in accordance with the provision of the protocol data transmission.The remote server achieved collecting,validating,displaying and saving the images based on the Netty framework.The test results showed that the transmission of BMP and the compressed JPEG images could meet the operational requirements of automated rice sowing test line.The transmission rate of JPEG images was greatly improved.The embedded data acquisition terminal could collect stable seeding tray images,and successfully uploaded to the server.The network average packet loss was 0.23% and the error rate was 0.23%.更多还原