【作者】 桂江生；

【导师】 应义斌；

【作者基本信息】 浙江大学 ， 农业机械化工程， 2007， 博士

【摘要】 在水果品质检测和分级中，形状是一个非常重要的指标，在国家标准中有严格的规定。本文在大量实验研究的基础上，通过对多种算法的对比分析，设计了具有统一架构的水果形状检测算法；从数学角度提出了形状的定义，并针对正常果形、轻度畸形、严重畸形三种类别研究了不同的形状描述方法；讨论了常用的分类器对形状分类结果的影响；最后针对基于形状描述的分类精度不是很高的情况，定义了新的测度并建立了基于配准技术的形状分类框架，达到了预期的研究目标。主要研究内容和研究结果如下：1．解决了如何从众多的去噪方法中选择适合水果形状检测的方法的难题。对于高斯噪声和随机噪声类型图像，在原始图像信噪比大于8时，选择全变差(Total Variation，TV)方法恢复可以达到最好效果；信噪比小于8时，选择维纳滤波才可以达到最好的恢复效果；而对于椒盐噪声类型图像，选用中值滤波时恢复效果最好。2．提出了运用矩阵广义逆和奇异值分解的方法恢复运动模糊图像，并用已知大小的标准球做实验，检测恢复后图像中球的大小，实验结果表明，提出的方法恢复后的图像质量要高于传统的盲去卷积等方法，速度要比差分递推法快6倍，比投影迭代方法快60倍。3．提出了多尺度水平集形状检测方法，解决了传统的方法无论是检测算子还是梯度向量流无法检测表面含有丰富颜色特征的水果形状的问题。实验结果表明该方法具有一系列的优点：无需任何形状预处理操作；具有一定的光照适应性；能够平滑地检测表面含有丰富颜色特征的水果，非常适合水果形状检测。4．提出了多尺度能量分布形状描述方法，将形状轮廓序列看作一周期信号，从多分辨率分析角度来说，代表形状全局信息的主要能量分布在粗尺度上，而表示形状局部信息的次要能量分布在细尺度上，该方法对严重畸形的水果分类比较有效，实验结果表明其分类精度可以达到81.20％。该方法中基于最大期望方法确定起始点的方法，可以唯一确定起始点，这对解决形状描述中旋转不变性问题非常有效。5．系统地分析和比较了目前常用形状描述方法，提出了将小波矩方法用于水果形状描述，得出结论：在水果形状分类中，用具有对称性的小波基(如Morlet小波)和最近邻法分类准则时，正常果形、轻度畸形及严重畸形的分类准确率可以达到69.42％、80.47％及72.62％。6．分析了分类器对水果形状分类结果的影响，集中研究了线性判别函数、聚类分析、BP神经网络和支持向量机等四种分类器分别对三种不同的输入特征模式(傅立叶描述子特征、Zernike矩特征和小波矩特征)时的分类性能。得出结论：无论用什么分类器，小波矩特征模式作为输入模式时，都能得到最好的分类结果；其次对小波矩特征模式而言，三个聚类中心的聚类分析方法和支持向量机作为分类器都能得到尚为满意的分类精度。采用三个聚类中心的聚类分析方法时，正常果形、轻度畸形及严重畸形的分类正确率为86.21％、65.78％和85.71％；而采用支持向量机作为分类器时，分类正确率分别为：70％、83.56％和75％。7．建立了基于配准技术的水果形状分类框架。定义了新的测度，该测度基于面积差水平集表示原理，将待分类的形状与已知类别的形状进行配准以使得它们的不重合面积最小。并从基于最小测度配准和运动估计配准两个方面进行实验，实验结果表明基于配准的方法能达到更高的分类精度和令人满意的结果，正常果形的分类正确率91.20％，轻度畸形的分类正确率为85.88％，严重畸形分类正确率为83.34％。更多还原

【Abstract】 In the process of fruit quality inspection and fruit sorting, shape is a very important index, which isprescribed by national Standard. Based on the comparison analysis of different algorithms and a numberof experiments, a united framework algorithm for shape detection was developed; the concept of shapewas defined from the mathematic view; and difference shape descriptors were analyzed for normalshape, slight abnormal shape and serious abnormal shape respectively, the effects of differenct classifierto class result were discussed; At last, for correction ration of the classification by shape descriptors wasnot satisfied with request, a new metric was designed and a new shape classification framework basedon registration technique wasconstructed, and expected research object was achieved. Main contentsand results were listed as follows:1. A difficult problem of how to select appropriate method for fruit shape detection from many imagedenoise techniques was solved. When the noise type was "Guassian" or "Random", and SNR wasover 8, TV method can achieve the best resume result, when SNR was under 8, Winner filter canachieve the best resume result; when the noise type was "salt&pepper", median filter can achievethe best resume result.2. A method based on General Matrix Inverse (GMI) and Singular Value Decomposition (SVD) wasdeveloped. Some standard balls were used for the experiments to detect the size of balls in theresumed image. The results showed that the new method was superior to the traditional blindconvolution and the speed is faster than difference successive 6 times and than project iteration 60times.3. Multi-scale level set algorithm for shape detection was developed, and resolved the problem thatfruit shape fruit shape which surface contain rich color features could not be detected by traditionalalgorithms not only detection operator but also Gradient Vector Flow. The experiment resultdemonstrated that this algorithm has many virtues: no need of any shape preprocess; is illuminationadaptive; can smoothly detect fruit shape which surface contain rich color features, So it is very fitfor fruit shape detection.4. A shape descriptors based on multi-scale energy distribution was proposed. The main idea was thatshape contour series could be viewed as one periodic signal, and from the view of multi-resolutionanalysis, the main energies representing global shape information were distributed at coarse scale,while the subordinate energies representing local shape information were distributed at fine scale.This method was effective for the classification serious abnormal fruit and experiment resultdemonstrated that the correct ratio can achieve 81.20%, the method for selection start point whichwas base on maximum expectation in this algorithm, could determine only start point, with wasvery effective to solute the invariant of rotation for shape description.5. The shape descriptors currently used were system analyzed, and wavelet moment was proposed forfruit shape description, and some conclusions were concluded that when wavelet base which is symmetry (e.g. Morlet) and nearest neighbor were used for fruit shape classification, the correctratio of normal shape, slight abnormal and serious abnormal can arrived at 69.42％, 80.47％and72.62％respectively.6. The influence of different classifiers on fruit shade classification was analyzed, especially on theinfluence of four classifiers (linear discrimination function, cluster analysis, back propagationneural network and support vector machine) combined with three feature patterns (Fourierdescriptor feature patterns, Zernike moment feature patterns and wavelet moment feature patterns).The conclusions were listed as follows: 1) best results were obtained by the wavelet featurepatterns combined with each classifier; 2) for wavelet moment feature patterns, good precision ofclassification were obtained by the classifier of cluster analysis (three cluster centre) and supportvector machine. The correction ration of normal shape, slight abnormality and serious abnormalitywas 86.21％, 65.78％and 85.71％when cluster analysis (three cluster center) was used. Whensupport vector machine was used as classifier, the correction ration of normal shape, slightabnormality and serious abnormality was 70％, 83.56％and 75％, respectively.7. A new method for shape classification based on registration was developed. In this method, a newmetric based on the principle of level set representation of difference area was designed. Minimalmetric registration and motion estimating registration were analyzed, the normal classificationcorrection ratio is 91.20％, slight abnormal classification correction ratio is 85.88％and seriousabnormal classification correction ratio was 83.34％. The result demonstrated that this methodbased on registration could obtain better classification correction ration than the traditional methodbased on shape description.更多还原