【作者】 王福杰；

【导师】 应义斌；

【作者基本信息】 浙江大学 ， 生物系统工程， 2013， 博士

【摘要】 果形是评价水果品质的重要特征之一。然而,由于自然形态差异大,水果形状的准确描述十分困难。目前形状描述方法存在执行复杂,对水果姿态要求严格,检测不全面等问题,不能满足高速分级过程中水果姿态多变情况下果形的描述。本文以苹果为主要研究对象,分别从二维和三维图像方面研究了水果形状的描述方法。研究了不同品种苹果的形状特征和单一品种不同等级苹果的形状特征的检测方法；构建了基于3相机的视觉检测系统,实现了姿态多变的运动水果的形状快速准确描述,确保姿态多变情况下水果形状判别的一致性。本文的主要研究内容、结果、结论如下(1)提出了一种FOFS背景分割与目标提取方法。针对水果颜色、大小、缺陷以及光照等多种因素变化导致OTSU等传统图像分割方法不能准确分割水果图像的问题,通过彩色图像的颜色分量融合、形态学开运算消噪、空间滤波去除锯齿状边界和自动阈值分割等步骤,准确地提取了目标。采用该方法对280幅苹果图像处理的结果表明：203幅图像分割偏差小于1%,占总量的72.50%；70幅图像分割偏差在1%-2%,占总量的25%；7幅图像偏差大于2%,占总量的2.50%；最大分割偏差为2.83%。结果表明,FOFS方法提取的目标偏差小于2.83%,边界光滑,具有良好的光照适应性,并可以较好地消除颜色、姿态、大小以及缺陷和果梗花萼对水果图像背景分割的影响。(2)提出了形心距离差指标的水果三维形状描述方法。针对水果二维形状描述方法只考虑水果的一个成像平面,形状描述不全面,精度差的问题,采用结构光三维成像技术采集水果的三维点云图像,根据形心距离差描述水果形状,对480幅三维点云图像进行处理的结果表明,对于具有端正果形的各品种苹果,形心距离差的描述方法可以100%区分细长形、近细长形、方圆形3种形状。结果表明,形心距离差三维形状描述方法可以描述水果细长形、近细长形、方圆形3种形状特征。(3)提出了对称指数、方形度、轮廓不平度、轮廓峭度、离心率等形状描述指标量,改进了傅里叶描述子、小波描述子以及各种矩的形状描述子。针对传统的形状特征量计算偏差大,不准确等问题,在对比分析了传统的描述方法的基础上,提出和改进了形状描述指标,对200个姿态多变的运动水果10次的形状检测试验证明,采用对称指数Os的离散度阈值区分形状端正果和畸变果的准确率达80.15%,采用改进的Zernike矩Z2-0阈值区分的准确率达66%以上；对440幅苹果图像处理的结果表明,采用Zernike矩Z3-1检测圆柱果和球形果的准确率达83.75%；对34组圆柱形和58组类球形,采用小波矩W3-8-1判定的准确率达93.10%。结果表明,对称指数Os、改进的Zernike矩、小波矩等形状描述子对检测水果形状具有一定的可行性。(4)提出了采用Zernike矩Z3-1和小波矩W3-8-1的阈值区分圆柱形果和类球形果的方法。针对目前形状检测主要关注的是单品种不同等级形状特征的描述,而不同品种的水果形状特征描述问题很少学者研究,本文以11个品种的苹果为研究对象,对比分析了244个形状描述指标,包括圆形度、占有率、对称指数、方形度、轮廓不平度、轮廓峭度、离心率、圆度、圆形比、内外接圆半径比,以及傅里叶描述子、小波描述子、Hu矩描述子、Zernike矩描述子和小波矩描述子。对440幅苹果图像处理的结果表明,采用Zernike矩Z3-1检测圆柱果和球形果的准确率达83.75%；对34组圆柱形和58组类球形,采用小波矩W3-8-1判定的准确率达93.10%。而采用多元线性回归模型,其判断准确率最高为56.52%。结果表明,采用Zernike矩Z3-1指标和小波矩W3-8-1指标的固定阈值对圆柱形和球形有一定的判别力,且比采用多元线性回归建立的模型分类准确率高,计算复杂度低。(5)提出了采用Zernike矩Z2-0的阈值判定多姿态水果形状是否端正,以及采用离心率Ec、圆形度C4、对称指数Os、Zernike矩Z5-3辅助判定特定姿态的水果形状是否端正的方法。针对目前形状描述方法主要研究的是特定姿态的水果形状,而多姿态的水果形状很少关注的问题,本文以采集于山东栖霞果园的29个富士苹果和36个乔纳金苹果为实验对象,根据形状人工将它们分成3个等级,对每个苹果采集了8幅不同姿态的图像。对比分析了244个形状描述指标,包括圆形度、占有率、对称指数、方形度、轮廓不平度、轮廓峭度、离心率、圆度、圆形比、内外接圆半径比,以及傅里叶描述子、小波描述子、Hu矩描述子、Zernike矩描述子和小波矩描述子。对520幅苹果图像处理的结果表明,采用Zernike矩Z2-0的阈值判别多姿态水果果形是否端正,对富士苹果和乔纳金苹果检测的准确率分别为67.24%和43.06%,而采用6种特征指标通过多元线性回归建立的线性模型判别,其对富士苹果和乔纳金苹果检测的准确率分别为11.21%和30.90%；果轴倾斜姿态,采用离心率Ec的阈值判别,其检测准确率对于富士苹果和乔纳金苹果分别为87.93%和79.17%。采用圆形度C4的阈值判别,其检测准确率对于富士苹果和乔纳金苹果分别为81.03%和65.28%。果轴竖直姿态,采用离心率Ec的阈值判别,其检测准确率对于富士苹果和乔纳金苹果分别为67.24%和59.72%。采用圆形度C4的阈值判别,其检测准确率对于富士苹果和乔纳金苹果分别为65.52%和59.72%。果轴水平姿态,采用Zernike矩Z5-3的阈值判别,其检测准确率对于富士苹果和乔纳金苹果分别为70.69%和52.08%。采用对称指数Os的阈值判别,其检测准确率对于富士苹果和乔纳金苹果分别为73.28%和59.03%。结果说明,根据Zernike矩Z2-0的阈值对多姿态苹果果形是否端正的判别有一定的可行性,且比采用多元线性回归建立的模型分类准确率高,计算复杂度低。在果轴竖直或者倾斜时,采用离心率Ec或者圆形度C4判别会更为准确；果轴水平时,采用Zernike矩Z5-3和对称指数Os对水果形状会更为准确。(6)设计了基于3相机的视觉系统,提出了采用Zernike矩Z2-0的阈值以及圆形度C4和对称指数Os的离散度阈值判定姿态多变运动水果的形状是否端正的方法。针对目前形状检测主要研究静止的固定姿态的水果,而姿态多变运动水果的形状检测一直是个难点。本文以水果市场购买的大小规格为90mm,80mm,70mm的三种红富士苹果为研究对象,对5个端正果,8个形变果,50次在线检测的1950幅图像处理的结果表明,采用Zernike矩Z2-0的阈值判别姿态多变运动水果的形状是否端正,5个形状端正果中有3个50次测试的检测结果都为端正果,其余2个50次均为错判;8个形变果中有5个50次测试的检测结果都为形变果,其余为部分错判。采用C4和Os的离散度判别,当C4的离散度小于10%时,判断其形状端正；当C4的离散度大于10%时,Os的离散度小于20%的判断其形状端正,该判定方法可以完全区分样本形状是否端正。进一步的验证实验,对100个端正果和100个形变果,10次在线检测的18000幅图像处理结果表明,采用3相机水果形状检测系统,设置采集一个工位的苹果3幅图像,根据这3幅图像的Zernike矩Z2-0的平均值的阈值判别水果形状,准确率约66%,平均每个苹果的检测时间约0.2063秒。设置采集一个苹果3个位置的9幅图像,根据这9幅图像的对称指数Os的离散度的阈值判别水果形状,准确率约80.15%,平均每个苹果的检测时间约1.649秒。结果表明,基于3相机视觉系统的水果形状检测,Zernike矩Z2-0阈值法和对称指数Os的离散度阈值法对果形是否端正具有一定的判别能力。采用Zernike矩Z2-0阈值法实时性好,采用对称指数Os的离散度阈值法准确率高。更多还原

【Abstract】 Shape is one of the key quality characteritcs of fruits. But, it is very difficult to describe the fruit shape beacause of its big difference among natural morphological characteristics of different fruits. The existing shape description methods mostly are facing many problems: complex executing, special fruits’postures, unilateral detection and low accuracy. They are not suitable for shape detection of pose-varied fruits on the high speed production line.In this paper, apples are mainly research objects. Shape description methods based on machine vision were studied from2D and3D images. The description methods of shape characteristics among different varieties of apples were studied, and the description methods of shape characteristics among different grades of a variety of apples were researched. A three cameras machine vision detection system was constructed, and fast and exact description of shape of pose-varied moving fruit was realized, and consistency of fruit shape discriminant under changing posture circumstances was guaranteed.The main research contents, results and conclusions were listed as follows:(1) A novel background segmentaion and target extraction method called FOFS was proposed. Using OTSU method directly cannot get a complete target because of influence of the fruit color, size, defects and illumination. In order to improve the background segmentation precision and adaptability to different fruits, the R, G, B companents of one color image were fusioned firstly, then the fusional image was denoised by morphological opening method, then followed by removing zigzag boundary through the spatial filtering, finally the background was segmented and the target was extracted by automatic threshold segmentation method.280apple images were processed by this method, and the results show that203image segmentation deviation is less than1%, accounting for72.50%of the total;70image segmentation deviation in1%～1%, accounting for25%of the total;7image deviation is greater than2%, accounting for2.50%of the total; the largest segmentation deviation is2.83%. The results indicate that the FOFS method is good. Segmentation deviation is less than2.83%, and the extraction edge is smooth, and it has good illumination adaptability. The FOFS method can better eliminate the influence of color, position, size, defects and stem calyx to background segmentation.(2) A3D shape description method of centroid distance deviation was presented. A2D image is only one of imaging plane of the apple, so shape detection by the description methods based on2D images was incomplete, and its accuracy was low.480images of3D points clouds of10varieties of apples by structured light3D imaging technology were analyzed by the method of centroid distance deviation. The results show that, for regular shape apples, the method of centroid distance deviation can distinguish slender, near slender, square and round. This results indicate that centroid distance deviation can describe shape features of slender, near slender, square and round.(3) Two dimensional shape description methods of symmetry index, rectangularity, contour irregularities, contour kurtosis and eccentricity were presented. Fourier descriptors, wavelet descriptors, Hu descriptors, Zernike moments descriptors,wavelet moment descriptors were improved. The traditional shape description indexes are inaccurate and they often have big deviation.so some new indexes were presented and some were improved by contrastive analysis with traditional indexes. The different shape grades of the same apple variety were detected by this method. The results showed that ten times of shape online detection of200apples by dispersion threshold of symmetry index Os proved the accuracy was about80.15%. And the distinguishing accuracy of the method of the improved Zernike moment Z2-0threshold was above66%. The results of the440apple images processing showed that the distinguishing accuracy of the method of the improved Zernike moment Z3-1threshold was above83.75%. The results of the34cylindrical apples and58spherical apples images processing showed that the distinguishing accuracy of the method of the improved wavelet moment W3-8-1threshold was above93.10%. This results indicate that symmetry index Os, improved Zernike moment and wavelet moment are useful.(4) A method of using Zernike moment Z3-1and Wavelet moment W3-8-1’s thresholds to classify the shape grade were proposed. In this paper,11varieties of apples were as research objects.244shape description indexes including circularity factor, complexity, occupancy, symmetry index, rectangularity, contour irregularities, contour kurtosis, eccentricity, roundness, circular ratio, the ratio of circumradius and Fourier descriptors, wavelet descriptors, Hu descriptors, Zernike moments descriptors, wavelet moment descriptors were analyzed. The results of440images processing show that the determining accuracy is above83.75%by setting Z3-1’s threshold12. greater than12cylindrical fruit, less than12spherical fruit. The determining accuracy is above93.10%to34groups of cylindrical and58groups of spherical fruits according to setting W3-8-1’s threshold5.5, greater than5.5clylindrical, less than5.5round. And using multiple linear regression model, its accuracy is56.52%. The results indicate that the Zernike moment Z3-1and Wavelet moment W3-8-1threshold method can recognize the shape of cylindrical and spherical. Its accuracy is higher than multiple liner regression model, and it has low computational complexity.(5) A method of using Zernike moment Z2-0’s threshold to judge deformation for all kinds of apple attitudes, and the method of eccentricity Ec, circularity factor C4, symmetry index Os, Zernike moment Z5-3’s thresholds for special attitudes apple images were proposed. Most researchers are focusing on shape description of special attitude fruit, and few care about shape description of multitude pose fruits. In this paper,29Fuji apples and36Jonah gold apples were picked from Qixia’ orchards in Shandong province. They were respectively selected to three shape grades according to their shape features by naked eyes. Each apple had8images from different attitudes.244shape description indexes including circularity factor, complexity, occupancy, symmetry index, rectangularity, contour irregularities, contour kurtosis, eccentricity, roundness, circular ratio, the ratio of circumradius and Fourier descriptors, wavelet descriptors, Hu descriptors, Zernike moments descriptors, wavelet moment descriptors were analyzed. The results of440images processing show that whatever any pose of apples, setting Z2-0’s threshold12, greater than12for distortionless, less than12for deformation, the detecting accuracy was about67.24%for Fuji apples and about43.06%for Jonah golden apples. And the detecting accuracy was only11.21%for Fuji apples and about30.09%for Jonah golden apples by multiple linear regression model. When fruit axis was leaned, the detecting accuracy of setting Ec’s threshold0.1was87.93%for Fuji apples and79.17%for Jonah golden apples. The detecting accuracy of setting C4’s threshold0.92was81.03%for Fuji apples and65.28%for Jonah golden apples. When fruit axis was vertical, the detecting accuracy of setting Ec’s threshold0.1was67.24%for Fuji apples and59.72%for Jonah golden apples. The detecting accuracy of setting C4’s threshold0.92was65.52%for Fuji apples and59.72%for Jonah golden apples. When fruit axis was horizonal, setting Os’s threshold0.9, greater than0.9for distortionless, less than0.9for deformation, the detecting accuracy was about73.28%for Fuji apples and about59.03%for Jonah golden apples. Setting Z5-3’s threshold11.9, greater than11.5for distortionless, less than11.5for deformation, the detecting accuracy was about70.69%for Fuji apples and about52.08%for Jonah golden apples. The results indicate that Zernike moment Z2-0can determine deformation. When fruit axis was vertical or leaned, eccentricity Ec and circularity factor C4can get better discrimination. When fruit axis was horizonal, symmetry index Os and Zernike moment Z5-3can get better discrimination. (6) A three camaras machine vision system was designed, and a method of using Zernike moment Z2-0’s threshold to judge deformation was proposed. And a method of using dispersion of circularity factor C4and symmetry index Os to judge deformation was presented. Most researchers are on shape description of fixed pose motionless fruit, and shape description of moving pose varied fruits is always difficult. From the fruits market, we bought Fuji apples that had3kinds of size specifications of90mm,80mm.70mm.5distortionless fruit and8deformational fruits were researched objects. And each apple must be tested for50times, and1950images were captured. According to the study on shape feature description methods of diffirent shape grades of the same apple variety, Zernike moment Z2-0, Z3-1, Z5-3and circularity factor C4, symmetry index Os, eccentricity Ec were analyzed mainly. The results show that setting Z2-0’s threshold13. greate than13for distortionless, less than13for deformation,3apples in5distortionless fruits were discriminated as distortionless fruits for50times tests and other2apples got the wrong results. And5apples in8deformational fruits were discriminated as deformation for50times tests and other apples had some wrong discrimination in some tests. The dispersion of circularity factor C4and symmetry index Os was calculated by statistic analysis method. Setting C4’s dispersion threshold10%, less than10%distortionless, when greater than10%. Os’s dispersion less than20%distortionless, other deformation, this method can completely discriminate the shape deformation or not. In validation tests,2kinds of distornless and deformation Fuji apples were selected. Each kind had100samples, and each sample must be tested for10times. And9imges of3postions for each time were captured, so18000images were obtained in total. The results of image processing show that, based on3camaras machine vision system, setting capturing3images from only1position, and, the accuracy of processing3images was about66%by Zernike moment Z2-0threshold12.7and average detection time for each apple was about0.2063seconds. Setting capturing9images from3positions, the accuracy of processing9images was about80.15%by dispersion threshold of symmetry index Os and average detection time for each apple was about1.649seconds. The results indicate that the methods of Zernike moment Z2-0threshold and dispersion threshold of symmetry index Os based on3camaras machine vision can discriminate the shape. The Zernike moment Z2-0threshold method has good real-time and the method of dispersion threshold of symmetry index Os has high accuracy.更多还原