基于I-RELIEF和SVM的畸形马铃薯在线分选

引用本文

张保华, 黄文倩, 李江波, 赵春江, 刘成良, 黄丹枫. 基于I-RELIEF和SVM的畸形马铃薯在线分选. 44(6): 1811-1817
ZHANG Bao-hua, HUANG Wen-qian, LI Jiang-bo, ZHAO Chun-jiang, LIU Cheng-liang, HUANG Dan-feng. Online sorting of irregular potatoes based on I-RELIEF and SVM method. Journal of Jilin University Engineering and Technology Edition, 44(6): 1811-1817 复制到剪切板

Permissions

基于I-RELIEF和SVM的畸形马铃薯在线分选

张保华^1,², 黄文倩², 李江波², 赵春江^1,², 刘成良¹, 黄丹枫¹

1.上海交通大学机械系统与振动国家重点实验室,上海 200240

2. 北京农业智能装备技术研究中心,北京 100097

通信作者:赵春江(1964-),男,教授,博士生导师.研究方向:农业机械化与自动化,农业信息化,E-mail:zhaocj@nercita.org.cn

作者简介:张保华(1986-),男,博士研究生.研究方向:基于机器视觉与模式识别的农产品无损检测.E-mail:zhangbaohua@sjtu.edu.cn

基金:国家自然科学青年基金项目(31301236); “863”国家高技术研究发展计划项目(2013AA100307); 2012年北京市农林科学院博士后基金项目;

摘要

在提取的RGB图像的R分量图像上提取二值图像和边界图像,计算马铃薯图像的偏心度、矩形度、圆形度以及边界图像的10个傅立叶描述子等共计13个形状特征。其次把样本的形状特征输入到I-RELIEF模块中,求出各个形状特征在分类中的影响程度,即各自权值。然后把带权值的样本形状特征输入到支持向量机算法模块中进行训练,从而得到分类器。最后用分类器实现畸形马铃薯的在线分选。试验结果表明:该方法每秒钟可以检测4个马铃薯,可以满足分选设备的实时性要求,并且畸形马铃薯的识别率高达98.1%。

关键词: 机器视觉; 马铃薯分选; 傅立叶描述子; 在线检测; 支持向量机

中图分类号:S126 文献标志码:A 文章编号:1671-5497(2014)06-1811-07

Online sorting of irregular potatoes based on I-RELIEF and SVM method

ZHANG Bao-hua^1,², HUANG Wen-qian², LI Jiang-bo², ZHAO Chun-jiang^1,², LIU Cheng-liang¹, HUANG Dan-feng¹

1.State Key Laboratory of Mechanical System and Vibration, Shanghai Jiaotong University, Shanghai 200240, China

2.Beijing Research Center of Intelligent Equipment for Agriculture, Beijing 100097, China

Abstract

An online sorting method of irregular potatoes is proposed. First, the R-component image is extract, and then the binary image and boundary image are obtained by thresholding method. Second, thirteen essential geometrical features, such as eccentricity, rectangle degree and roundness, and ten Fourier descriptors are extracted. Third, the geometrical features of the potato image sample are feed into the module of I-RELIEF algorithm, which exports a weight for each feature. Fourth, the features of the potato image sample with weights are feed into the training module of Support Vector Machine (SVM). Finally, the classifier model is used to make decision and achieve the grading result online based on the potato's features and weights. Results show that the SVM method can detect and sort four potatoes per second with the help of I-RELIEF module, and the overall accuracy is 98.1%.

Keyword: computer vision; potato sorting; fourier descriptor; in-line detection; support vector machine(SVM)

Show Figures

0 引言

机器视觉技术是一种客观、快速、经济的检测技术^{[ 1, 2]},将其应用于农产品品质检测具有巨大的潜在价值^{[ 3, 4, 5, 6, 7, 8, 9]}。马铃薯在形状和尺寸上波动范围较大,这使得马铃薯的检测和分选变得困难。在市场交易时,具有规则外形的马铃薯深受消费者喜爱,不规则的畸形马铃薯因外观怪异、食用时不便于削皮且有较大浪费等原因则遭到冷落。因此交易前利用机器视觉技术对具有不同外观形状的马铃薯进行分选是必要的。

Heinemann等^{[ 7]}利用边界提取算法提取马铃薯的边界图像,通过提取边界图像的傅立叶描述子作为形状特征,然后根据形状和尺寸对马铃薯进行分选。Zhou等^{[ 10]}开发了一个针对马铃薯质量、形状、切面直径、颜色等特征进行马铃薯分选的计算机视觉系统,分选准确率在85%以上。Navid等^{[ 11]}提出了基于机器视觉的马铃薯外部缺陷和形状的在线检测方法,形状特征选取了外接矩形长宽比、最大直径和最小直径3个特征,融合马铃薯的缺陷特征对马铃薯进行在线分选,结果证明支持向量机的方法在准确率和效率上具有较好的表现,分选准确率在95%以上。Gamal等^{[ 12]}提出了一种基于机器视觉的马铃薯形状实时检测方法,该方法提取了离心率、偏心度、圆形度、伸长量4个基于傅立叶的相关参数,对马铃薯进行在线分选,总分选精度为96%。内蒙古农业大学郝敏等^{[ 13]}提取马铃薯的Zernike矩作为特征参数并利用支持向量机实现了马铃薯外形的分类。在对马铃薯形状进行分类时,现有技术大多没有考虑各个形状特征对分类的影响程度,相关性较小的特征不仅对分类准确率没有太多贡献,而且还浪费了特征提取所需的时间,不利于马铃薯快速在线分选。

为了满足马铃薯分选的实时性要求,本文在离线训练阶段,初步选用了马铃薯的偏心度、矩形度、圆形度和边界图像的10个傅立叶描述子作为马铃薯的形状特征,然后基于I-RELIEF算法对所选特征进行权值计算,剔除相关性较小的特征,同时记录所选形状特征的权值向量,最后把形状特征和权值向量一并送入支持向量机进行训练,生成马铃薯形状分类器。在在线检测阶段,在线提取马铃薯的形状特征,并加权处理,利用离线生成的分类器依据形状完成马铃薯的分选。

1 材料及方法

1.1 试验材料

试验用马铃薯样本购买于北京当地菜市场,共计282个,通过后期筛选,选择其中260个作为试验样本,样本分为正常马铃薯和畸形马铃薯,其中正常马铃薯共有170个,畸形马铃薯共有90个。正常马铃薯一般呈现类圆形或者椭圆形,畸形马铃薯多呈现马铃薯块茎含有畸形凸起、不规则歪曲,或由于马铃薯块茎二次发育造成主块茎附带一个或多个小块茎^{[ 14, 15]}。

1.2 马铃薯自动分选系统

马铃薯自动分选系统如图1所示,主要包括上料装置、传输装置、控制装置、计算机视觉检测系统、卸料装置等部分。

	Figure Option View Download New Window
	图1 马铃薯自动分选系统示意图Fig.1 Diagram of general scheme

系统的工作流程为:系统上料,马铃薯单列置于托盘上随传送链条传动。光电编码器检测到马铃薯,信号传输至计算机并控制相机采集马铃薯图像,经过畸形马铃薯检测软件,对马铃薯的外观进行检测、分类。控制系统控制卸料装置卸料,完成马铃薯的分选。

1.3 本文算法

首先,通过摄像机获取马铃薯的RGB图像,分离得到R通道图像,随后,采用亮度校正算法对R通道图像进行亮度校正,校正算法如下:

$\begin{matrix} Output = 255 \times \frac{Input - Min}{Max - Min} (1) \end{matrix}$

式中: Max、 Min为R通道图像中的最大、最小灰度值; Input为R通道图像的像素灰度值; Output为变换后对应坐标点的像素灰度值。

在获取亮度校正图像后,首先对图像采用3×3圆盘形结构元素进行高斯滤波,然后使用自适应阈值完成马铃薯图像的分割。为了消除二值图像中存在的小区域噪声和马铃薯瓜蒂部位的细线状噪声,采用区域填充算法和形态学开运算对二值图像进行操作。然后,对马铃薯二值图像进行分析,提取马铃薯的偏心率、矩形度、圆形度和10个傅立叶描述子等形状特征。最后,将所选特征和其权值向量一并送入支持向量机模型进行形状分类。

1.4 分类器设计流程图

在马铃薯形状分类过程中,每一个形状特征对马铃薯形状类别识别的作用程度并不相同,一些相关性较小的形状特征的参与,不利于马铃薯形状的快速准确检测。因此本文在训练分类器之前,运用了I-RELIEF算法对初步选择的形状特征进行权值计算,然后剔除相关性较小的特征,并记录所选用特征的权值向量,在离线训练阶段和在线预测阶段,马铃薯的形状特征向量和权值向量一并送入分类器进行分类。图2为本文分类器设计的流程图。

	Figure Option View Download New Window
	图2 分类器设计流程图Fig.2 Flowchart of the training and predicting classifier

2 马铃薯形状特征的选择和提取

(1)偏心率。偏心率也称为伸长度,在一定程度上反映了区域的紧凑性。物体的偏心率定义为最小外接矩形的主轴意义下的长宽比,公式为:

$\begin{matrix} E = \frac{A}{B} (2) \end{matrix}$

式中: A为马铃薯区域最小外接矩形主轴长度; B为马铃薯区域最小外接矩形辅轴长度。

马铃薯的最小外接矩形以及其主轴和辅轴如图3所示。

	Figure Option View Download New Window
	图3 主轴和辅轴Fig.3 Main and auxiliary axis

(2)矩形度。矩形度定义为图像中马铃薯的面积和其最小外接矩形面积的比值,矩形度体现了马铃薯对其最小外接矩形的充满程度,因此矩形度大于0小于1,其计算公式为:

$\begin{matrix} R = \frac{A_{O}}{A_{MER}} (3) \end{matrix}$

式中: A_O表示马铃薯的像素面积; A_MER表示马铃薯最小外接矩形的面积。

马铃薯的面积以及马铃薯最小外接矩形面积如图4所示。

	Figure Option View Download New Window
	图4 矩形度求取方法Fig.4 Area of potato and it’s minimum enclosing rectangle

正常的马铃薯对其最小外接矩形的充满程度较大,因此矩形度较大;而畸形马铃薯对其最小外接矩形的充满程度较小,因此矩形度较小。所以把矩形度作为一个衡量马铃薯形状的特征纳入特征集。

(3)圆形度。圆形度是一个用来刻画物体边界复杂程度的量。其计算公式为:

$\begin{matrix} e = \frac{4 π A_{O}}{L^{2}} (4) \end{matrix}$

式中: L表示马铃薯二值图像的周长,其中,4连通中相邻像素距离为1,而8连通中斜对角相邻像素距离为 $\begin{matrix} \sqrt[]{2} \end{matrix}$ 。

接近圆形的马铃薯的外形比较简单,圆形度较大;畸形马铃薯外形复杂,圆形度较小。因此圆形度在一定程度上反映了马铃薯的外形特征,被纳入马铃薯形状的特征集。

(4)傅立叶描述子。在图像中,物体的边界是一个难以定量描述的二维离散曲线段,对边界曲线段作离散傅立叶变换,是定量描述边界形状的一种常用方式。傅立叶变换可以将 x-y平面中的二维曲线段转换成复平面上的一个序列,对该序列进行一维的离散傅立叶变换,从而得到一系列的傅立叶系数,这个系数便是边界的傅立叶描述子。在现实中,直接把空间平面和复平面重合,空间平面上的点和其在复平面上的点一一对应,这样就完成了从空域到频域的坐标转化。坐标转换关系如图5所示,其中横坐标表示空间 x轴或复平面 u轴,纵坐标表示空间 y轴或复平面 v轴;( x_k, y_k)代表边界上任一点, u_k+jv_k表示复平面中与其对应的点。

	Figure Option View Download New Window
	图5 空间域和频域坐标转换关系Fig.5 Conversion between space domain and frequency domain coordinates

假设一个封闭的边界曲线段由 N点组成,其傅立叶描述子可以通过如下方法求取:

(1)将 x-y平面和复平面 u-v坐标系重合,则构成边界的点的坐标在空域和频域一一对应,从任一点绕边界一周,得到边界的复数序列:

$\begin{matrix} \begin{matrix} s (k) = u (k) + jv (k) (5) \\ k = 0,1, \dots, N - 1 \end{matrix} \end{matrix}$

(2)对序列 s( k)作离散傅立叶变换,得到边界的傅立叶描述:

$\begin{matrix} \begin{matrix} S (ω) = \frac{1}{N} \overset{N - 1}{\sum_{k = 0}} S (k) \exp (- j \frac{2 π ωk}{N}) (6) \\ ω = 0,1, \dots, N - 1 \end{matrix} \end{matrix}$

(3)对傅立叶描述进行逆变换:

$\begin{matrix} \begin{matrix} S (k) = \frac{1}{N} \overset{N - 1}{\sum_{k = 0}} S (ω) \exp (- j \frac{2 πωk}{N}) (7) \\ k = 0,1, \dots, N - 1 \end{matrix} \end{matrix}$

(4)傅立叶变换的低频成分是对边界形状的描述,因此取前 M个系数即可以得到一个近似的表示:

$\begin{matrix} \begin{matrix} \bar{s} (k) = \frac{1}{N} \overset{M - 1}{\sum_{k = 0}} S (ω) \exp (- j \frac{2 πωk}{N}) (8) \\ k = 0,1, \dots, N - 1 \end{matrix} \end{matrix}$

傅立叶变换系数的模具有平移和旋转不变的特性,因此可用傅立叶变换的系数向量作为物体形状的定量描述。在实际运用中,通常将其幅值归一化,如除以其平均幅值或最大幅值,可以得到与尺度无关的形状特征。本文把傅立叶描述子的前10个变换系数作为描述马铃薯形状的特征集。

3 I-RELIEF和支持向量机算法

3.1 I-RELIEF算法

本文选择了13个描述马铃薯外形的特征,把这些对分类器发挥的作用相差较大的特征一视同仁地作为支持向量机的输入是不合理的。在特征选择算法中,RELIEF是一种简单有效的方法,然而RELIEF算法对于调和参数选择非常敏感,鲁棒性较差。Sun等^{[ 16]}为了克服上述缺点并且降低分类算法的时间消耗,对RELIEF方法进行了改进,提出了I-RELIEF算法,本文运用了该算法对马铃薯形状特征进行加权处理。

(1)I-RELIEF算法的概念:

Hit集: H_n={ i:1≤ i≤ N, l_i=l_n, i≠ n},此集合为样本集中与样本 x_n属于相同类别的样本序号集。

Miss集: H_n={ i:1≤ i≤ N, l_i≠ l_n},此集合为样本集中与样本 x_n属于不同类别的样本序号集。

(2)RELIEF算法的目标函数为:

$\begin{matrix} \begin{matrix} C (w) = \\ \frac{1}{N} \overset{N}{\sum_{n = 0, o_{n} = 1}} (‖ x_{n} - x_{S_{n 1}} ‖_{w} - ‖ x_{n} - x_{S_{n 2}} ‖_{w}) (9) \end{matrix} \end{matrix}$

I-RELIEF算法就是为了优化目标函数(见式(9))而提出来的改进方法,经过优化计算,I-RELIEF算法获得权向量 w。其中, S_n₁为Miss集中与样本 x_n距离最近的样本序号; S_n₂为Hit集中与样本 x_n距离最近的样本序号;‖ x‖ _w= $\begin{matrix} \overset{N}{\sum_{i = 1}} \end{matrix}$ w_i‖ x_i‖; o=[ o₁, o₂,…, o_n]^T为参数向量;当 x_n为正常样本时 o_n=1,异常样本时 o_n=0。

样本 x_i是样本集 x_n中最接近Miss样本的概率为:

$\begin{matrix} \begin{matrix} α_{i, n} = P_{m} (i |x_{n}, w^{(t)}) = \\ \frac{f (‖ x_{n} - x_{i} ‖_{w})}{\sum_{{j\in M_{n}}_{}}^{_{}} f (‖ x_{n} - x_{i} ‖_{w})}, \forall i \in M_{n} (10) \end{matrix} \end{matrix}$

样本 x_i是样本集 x_n中最接近Hit样本的概率为:

$\begin{matrix} \begin{matrix} β_{i, n} = P_{m} (i |x_{n}, w^{(t)}) = \\ \frac{f (‖ x_{n} - x_{i} ‖_{w})}{\sum_{{j\in H_{n}}_{}}^{_{}} f (‖ x_{n} - x_{i} ‖_{w})}, \forall i \in H_{n} (11) \end{matrix} \end{matrix}$

样本属于异常样本的概率为:

$\begin{matrix} \begin{matrix} γ_{n} = P_{O} (o_{n} = 0 |D, w^{(t)}) = \\ \frac{\sum_{{j\in M_{n}}_{}}^{_{}} f (‖ x_{n} - x_{i} ‖_{w})}{\sum_{{j\in H_{n}}_{}}^{_{}} f (‖ x_{n} - x_{i} ‖_{w})} (12) \end{matrix} \end{matrix}$

式中: f(·)为各式的核函数。

(3)I-RELIEF算法的实现方法:

为了优化目标函数 C( w),需要如下计算迭代式:

$\begin{matrix} \begin{matrix} Q (w |w^{(t)}) = E_{\{S, O\}} [C (w)] = \\ w^{T} \frac{1}{N} \overset{N}{\sum_{n = 1}} γ_{n} ({\bar{m}}_{n} - {\bar{h}}_{n}) (13) \end{matrix} \end{matrix}$

式中: $\begin{matrix} {\bar{m}}_{n} \end{matrix}$ = $\begin{matrix} \sum_{i \in M_{n}} \end{matrix}$ α_i_, _n $\begin{matrix} m_{i, n}^{j} \end{matrix}$ ; $\begin{matrix} {\bar{h}}_{m} \end{matrix}$ = $\begin{matrix} \sum_{i \in H_{n}} \end{matrix}$ β_i_, _n $\begin{matrix} m_{i, n}^{j} \end{matrix}$ 。

式(13)中的 Q( w $\begin{matrix} |w^{(t)} \end{matrix}$ )是第 t次迭代后的结果,记:

$\begin{matrix} v = \frac{1}{N} \overset{N}{\sum_{n = 1}} γ_{n} ({\bar{m}}_{n} - {\bar{h}}_{n}) (14) \end{matrix}$

对于第 t+1次迭代, w可由下式求出:

$\begin{matrix} w^{(t + 1)} = \arg \max_{w \in W} Q (w |w^{(t)}) = \frac{v}{‖ v ‖_{2}} (15) \end{matrix}$

式中: w={ w:‖ w $\begin{matrix} ‖_{2}^{2} \end{matrix}$ =1, w≥0}。

重复迭代式(14)和式(15),直至满足:

$\begin{matrix} ‖ w^{(t + 1)} - w^{(t)} ‖< θ (16) \end{matrix}$

式中: θ表示迭代终止条件。

通过以上算法可以得到特征集的权重系数向量 w,权重系数的大小反映了相关特征对马铃薯形状类别归属的作用程度,因此权重系数过小则可以作为无关特征予以去除。将训练集数据输入I-RELIEF算法模块中进行计算,将得出的权重系数向量作为支持向量机输入特征的权重向量,对在线马铃薯的形状进行分类。

3.2 支持向量机算法

支持向量机是由Corinna Cortes和Vapuik在1995年首次提出来的^{[ 17]}。该方法建立在统计学习理论的VC维理论和结构风险最小原理基础上,根据有限的样本信息在模型的复杂性和学习能力之间寻求最佳折衷,以期获得最好的推广能力。本文采用支持向量机作为分类器,选择径向基作为核函数,惩罚因子 C=10.0,核函数参数 γ=8.0。

4 试验结果及分析

本文把马铃薯分为畸形马铃薯和正常马铃薯两类,两种类别的马铃薯样本是由一位具有丰富经验的专业分选工人挑选的,且在样本选择上,尽量使样本覆盖几乎所有可能的马铃薯大小和形状级别。每类马铃薯分别选取50个作为训练样本,离线训练和在线预测阶段的马铃薯图像都是在分类生产线上在线获取的,每秒钟采集4幅,为了更加准确、全面地反映马铃薯的真实形态,本文在训练阶段对样本连续采集3次。

在离线训练阶段,获得图像后,首先对图像进行预处理、二值分割、边界跟踪;然后提取了马铃薯的偏心率、矩形度、圆形度和10个傅立叶描述子系数共计13个特征作为马铃薯的形状特征。图像处理过程如图6所示。

	Figure Option View Download New Window
	图6 马铃薯图像处理过程Fig.6 Steps of the processing of images

把这些样本以及样本类别标号输入到I-RELIEF算法模块中,计算出各个特征量的权值,在在线预测时,由这些权值构成权值向量,与特征向量一起送入支持向量机中进行分类预测。经过I-RELIEF算法的计算,各个特征的权值如图7所示, 图7中,1~13分别代表偏心率、矩形度、圆形度及傅里叶描述子的前10个系数。

	Figure Option View Download New Window
	图7 马铃薯形状特征权值Fig.7 Weights of the features

从图7可以看出:偏心率、圆形度、傅立叶描述子的第3个系数在分类中发挥的作用较大;而矩形度、傅立叶描述子第2个系数、第7个系数在分类时发挥的作用小,因此剔除马铃薯形状的这3个特征,把剩余权值组成权值向量,与其相对应的特征一起送入支持向量机分类器,然后得出分类器模型。

在线预测阶段,首先载入已经训练好的分类器模型;其次根据光电编码器的触发信号,抓取传送滚子上的马铃薯图像,对图像预处理、二值化、边界跟踪;然后提取马铃薯图像的偏心率、圆形度和10个傅立叶描述子特征;最后利用分类器对马铃薯的形状进行分类。

对120个正常马铃薯和40个畸形马铃薯进行在线预测,为了使预测数据更具合理性和可靠性,本文对正常马铃薯和畸形马铃薯进行混合试验,通过上料机构随机进行检测,且重复进行3次独立试验,综合分析计算了检测正确率。分选结果如表1所示。正常马铃薯有6个被识别为畸形马铃薯,准确率达到98.3%。被误识别的马铃薯有2个是因为马铃薯细长,且侧面图像被采集,另外有1个出现过分割的问题,有3个是因为马铃薯被擦破表皮所致。畸形马铃薯有3个被误识别为正常马铃薯,准确率为97.5%。被误识别的马铃薯是因为在传送过程中,马铃薯旋转,采集的图像为其侧面,因而畸形没有被检测出来。虽然本文多次采集马铃薯的图像,但是每次都是独立进行检测的,为了提高精度,在以后的研究中需要设计更加合理的采集系统,例如可以使用多个相机系统采集不同方位的马铃薯的图像,多图像融合对马铃薯形状做出判断。

表1 试验结果 Table 1 Results of the experiment

5 结束语

开发了一种基于I-RELIEF和SVM的畸形马铃薯在线检测识别算法。该算法首先提取马铃薯图像的偏心率、矩形度、圆形度和10个傅立叶描述子共计13个特征。然后把这13个形状特征输入I-RELIEF算法模块,求出各个特征的权值。计算结果表明:偏心率、圆形度以及傅立叶描述子的第3个系数在分类中发挥的作用较大,而矩形度、傅立叶描述子第2个系数、第7个系数在分类时发挥的作用小,剔除这3个作用较小的特征;最后把所剩余的10个形状特征和其权值一起送入SVM算法中进行训练,得到马铃薯形状分类器。在马铃薯的在线分选阶段便可以载入分类器对马铃薯进行分选。试验结果表明,该算法每秒钟可以检测4个马铃薯,满足在线检测的实时性要求,并且畸形马铃薯的识别率达到98.1%。

The authors have declared that no competing interests exist.

参考文献

View Option

[1]	Moreda G P, Muñoz M A, Ruiz-Altisent M, et al. Shape determination of horticultural produce using two-dimensional computer vision—a review[J]. Journal of Food Engineering, 2012, 108(2): 245-261. [本文引用:1] [JCR: 2.276]
[2]	Zhang B H, Huang W Q, Li J B, et al. Principles, development and applications of computer vision for external quality inspection of fruits and vegetables: A review[J]. Food Research International, 2014, 62: 326-343. [本文引用:1] [JCR: 3.005]
[3]	Tao Y, Morrow C T, Heinemann P H, et al. Fourier-based separation technique for shape grading of potatoes using machine vision[J]. Transactions of the ASAE, 1995, 38(3): 949-957. [本文引用:1] [JCR: 0.822]
[4]	Corrado C, Francesca A, Federico P, et al. Shape analysis of agricultural products: a review of recent research advances and potential application to computer vision[J]. Food Bioprocess Technol, 2011, 4(5): 673-692. [本文引用:1]
[5]	Al-Mallahi A, Kataoka T, Okamoto H, et al. An image processing algorithm for detecting in-line potato tubers without singulation[J]. Computers and Electronics in Agriculture, 2010, 70(1): 239-244. [本文引用:1] [JCR: 1.766]
[6]	Gomes J F S, Leta F R. Application of computer vision techniques in the agriculture and food industry: a review[J]. European Food Research and Technology, 2012, 235(6): 989-1000. [本文引用:1] [JCR: 1.436]
[7]	Heinemann P H, Pathare N P, Morrow C T. An automated inspection station for machine-vision grading of potatoes[J]. Machine Vision and Applications, 1996, 9(1): 14-19. [本文引用:1] [JCR: 1.103]
[8]	Patel K K, Kar A, Jha S N, et al. Machine vision system: a tool for quality inspection of food and agricultural products[J]. Journal of Food Science and Technology, 2012, 49(2): 123-141. [本文引用:1]
[9]	Tao Y, Heinemann P H, Varghese Z. Machine vision for color inspection of potatoes and apples[J]. Transactions of the ASAE, 1995, 38(5): 1555-1561. [本文引用:1] [JCR: 0.822]
[10]	Zhou L, Chalana V, Kim Y. PC-based machine vision for real-time computer-aided potato inspection[J]. International Journal of Imaging Systems and Technology, 1998, 9(6): 423-433. [本文引用:1] [JCR: 0.639]
[11]	Navid R, Somayeh M B, Soleymani F. A real-time mathematical computer method for potato inspection using machine vision[J]. Computer and Mathematics with Application, 2012, 63(1): 268-279. [本文引用:1]
[12]	Gamal E, Sergio C, Enrique M, et al. In-line sorting of irregular potatoes by using automated computer-based machine vision system[J]. Journal of Food Engineering, 2012, 112(1-2): 60-68. [本文引用:1] [JCR: 2.276]
[13]	郝敏, 麻硕士, 郝小冬. 基于Zernike矩的马铃薯形状检测[J]. 农业工程学报, 2010, 26(2): 347-350. Hao Min, Ma Shuo-shi, Hao Xiao-dong. Potato shape detection based on Zernike moments[J]. Transactions of the Chinese Society of Agricultural Engineering, 2010, 26(2): 347-350. [本文引用:1] [CJCR: 1.299]
[14]	周竹, 黄懿, 李小昱, 等. 基于机器视觉的马铃薯自动分级方法[J]. 农业工程学报, 2012, 28(7): 178-183. Zhou Zhu, Huang Yi, Li Xiao-yu, et al. Automatic detecting and grading method of potatoes based on machine vision[J]. Transactions of the Chinese Society of Agricultural Engineering, 2012, 28(7): 178-183. [本文引用:1] [CJCR: 1.299]
[15]	郑冠楠, 谭豫之, 张俊雄, 等. 基于计算机视觉的马铃薯自动检测分级[J]. 农业机械学报, 2009, 40(4): 166-168, 156. Zheng Guan-nan, Tan Yu-zhi, Zhang Jun-xiong, et al. Automatic detecting and grading method of potatoes with computer vision[J]. Transactions of the CSAM, 2009, 40(4): 166-168, 156. [本文引用:1]
[16]	Sun Yi-jun. Iterative RELIEF for feature weighting: algorithms, theories, and application[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2007, 29(6): 1035-1051. [本文引用:1]
[17]	Cortes C, Vapuik V. Soft margin classifier: U. S. Patent 5640492[P]. 1997-6-17. [本文引用:1]

2012

2.276

0.0

. 2012, 108(2):null-null

Shape determination of horticultural produce using two-dimensional computer vision—a review

2014

3.005

0.0

. 2014, 62:null-null

Principles, development and applications of computer vision for external quality inspection of fruits and vegetables: A review

1995

0.822

0.0

2011

0.0

. 2011, 4(5):673-692 DOI:10.1007/s11947-011-0556-0

Shape analysis of agricultural products: a review of recent research advances and potential application to computer vision

1.Agricultural Engineering Research Unit of the Agriculture Research Council, CRA-ING, Via della Pascolare, 16, 00015, Monterotondo Scalo, Rome, Italy 2.Instituto de Ciencías del Mar (ICM-CSIC), Paseo Marítimo de la Barceloneta, 37-49, 08003 Barcelona, Spain 3.Food Refrigeration and Computerized Food Technology, Agriculture and Food Science Centre, University College Dublin, Belfield, Dublin, 4 Ireland

The appearance of agricultural products deeply conditions their marketing. Appearance is normally evaluated by considering size, shape, form, colour, freshness condition and finally the absence of visual defects. Among these features, the shape plays a crucial role. Description of agricultural product shape is often necessary in research fields for a range of different purposes, including the investigation of shape traits heritability for cultivar descriptions, plant variety or cultivar patents and evaluation of consumer decision performance. This review reports the main applications of shape analysis on agricultural products such as relationships between shape and: (1) genetic; (2) conformity and condition ratios; (3) products characterization; (4) product sorting and finally, (5) clone selection. Shape can be a protagonist of evaluation criteria only if an appreciable level of image shape processing and automation and data are treated with solid multivariate statistic. In this context, image-processing algorithms have been increasingly developed in the last decade in order to objectively measure the external features of agricultural products. Grading and sorting of agricultural products using machine vision in conjunction with pattern recognition techniques offers many advantages over the conventional optical or mechanical sorting devices. With this aims, we propose a new automated shape processing system which could be useful for both scientific and industrial purposes, forming the bases of a common language for the scientific community. We applied such a processing scheme to morphologically discriminate nuts fruit of different species. Operative Matlab codes for shape analysis are reported.

2010

1.766

0.0

. 2010, 70(1):239-244 DOI:10.1016/j.compag.2009.11.001

An image processing algorithm for detecting in-line potato tubers without singulation

Abstract

An algorithm which helps detecting in-line potato tubers without singulation when they are mixed with clods was developed. It is particularly useful for sorting and grading purposes whether at the packhouses or the potato harvester. The algorithm scans images which include clusters of tubers and clods from left to right until it encounters a cluster. Once encountered, a scan around the periphery of the cluster starts until returning to the original point where the scanning around the periphery started. At the end of the periphery scanning, the periphery pixels are erased and the new width of the cluster is compared with the previous width. Contouring the periphery continues until a sudden shrink in the width of the cluster occurs which indicates the detection of a contact point. The left-to-right scanning resumes from the contact point until it reaches the right end of the image repeating the periphery scanning every time a cluster is encountered.

The algorithm was tested by detecting 380 contact points between tubers and 112 contact points between tubers and clods. Results showed that approximately 99% of the contact points were detected successfully. It was found that the only reason of error was the lack of cusp between a tuber and a clod when the clod did not have an oval shape. It could be concluded that the algorithm is able to detect any contact point between any oval shaped adjacent objects as long as they form a cusp between them, which is guaranteed between any adjacent tubers. Therefore, the algorithm can be applied at the packhouses to replace the long, bulky and expensive conveyors needed to create spaces between the in-line potato tubers so as to aid machine vision-based size grading. Moreover, the algorithm can be used at the potato harvester to help in discriminating between the touching tubers and clods so as to automate the removal of the clods during harvesting.

2012

1.436

0.0

. 2012, 235(6):989-1000 DOI:10.1007/s00217-012-1844-2

Application of computer vision techniques in the agriculture and food industry: a review

<li>1.Colorimetry and Spectrophotometry Laboratory (Lacoe), Optical Metrology Division (Diopt), Brazilian National Institute of Metrology, Quality and Technology (Inmetro), Av. Nossa Senhora das Graças, 50, Building 4, Xerém, Duque de Caxias, RJ, CEP 25250-020, Brazil</li><li>2.Computational and Dimensional Metrology Laboratory (LMDC), Programa de Pós-Graduação em Engenharia Mecânica (PGMEC), Universidade Federal Fluminense (UFF), R. Passo da Pátria, 156, Bloco E, Campus da Praia Vermelha, São Domingos, Niterói, RJ, CEP 24210-240, Brazil</li>

Abstract
Over the last decades, parallel to technological development, there has been a great increase in the use of visual inspection systems. These systems have been widely implemented, particularly in the stage of inspection of product quality, as a means of replacing manual inspection conducted by humans. Much research has been published proposing the use of such tools in the processes of sorting and classification of food products. This paper presents a review of the main publications in the last ten years with respect to new technologies and to the wide application of systems of visual inspection in the sectors of precision farming and in the food industry.

1996

1.103

0.0

Machine Vis. Apps.. 1996, 9(1):14-19

An automated inspection station for machine-vision grading of potatoes

Paul H. Heinemann (1) , Niranjan P. Pathare (2) , Charles T. Morrow (1)

1.249 Agricultural Engineering Building, 16802, University Park, PA, USA 2.Department of Engineering, SGS-Thomson Microelectronics, M/S 736, 1310 Electronics Drive, 75006, Carrollton, TX, USA

A prototype inspection station based on the United States Department of Agriculture (USDA) inspection standards was developed for potato grading. The station consisted of an imaging chamber, conveyor, camera, sorting unit, and personal computer for image acquisition, analysis, and equipment control. A sample of 9.1kg (201b) of pregraded potatoes were evaluated in three separate experimental runs to assess the system performance. The system correctly classified 80%, 77%, and 88% of the moving potatoes in the three runs at 3 potatoes/min, and 98%, 97%, and 97%, in three runs of stationary potatoes. Shape analysis was adversely affected by the potato motion, and this contributed to the misclassification error.

... Heinemann等^[7]利用边界提取算法提取马铃薯的边界图像,通过提取边界图像的傅立叶描述子作为形状特征,然后根据形状和尺寸对马铃薯进行分选 ...

2012

0.0

. 2012, 49(2):123-141 DOI:10.1007/s13197-011-0321-4

Machine vision system: a tool for quality inspection of food and agricultural products

<li>1.Division of Post Harvest Technology, Indian Agricultural Research Institute, New Delhi, 110012, India</li><li>2.CIPHET, Ludhiana, Punjab, India</li><li>3.AMU, Aligarh, 202002, Uttar Pradesh, India</li>

Abstract
Quality inspection of food and agricultural produce are difficult and labor intensive. Simultaneously, with increased expectations for food products of high quality and safety standards, the need for accurate, fast and objective quality determination of these characteristics in food products continues to grow. However, these operations generally in India are manual which is costly as well as unreliable because human decision in identifying quality factors such as appearance, flavor, nutrient, texture, etc., is inconsistent, subjective and slow. Machine vision provides one alternative for an automated, non-destructive and cost-effective technique to accomplish these requirements. This inspection approach based on image analysis and processing has found a variety of different applications in the food industry. Considerable research has highlighted its potential for the inspection and grading of fruits and vegetables, grain quality and characteristic examination and quality evaluation of other food products like bakery products, pizza, cheese, and noodles etc. The objective of this paper is to provide in depth introduction of machine vision system, its components and recent work reported on food and agricultural produce.

1995

0.822

0.0

1998

0.639

0.0

. 1998, 9(6):423-433

PC-based machine vision system for real-time computer-aided potato inspection

Liyong Zhou,Vikram Chalana andYongmin Kim *

Image Computing Systems Laboratory, Department of Electrical Engineering, Box 352500, University of Washington, Seattle, WA 98195 *Image Computing Systems Laboratory, Department of Electrical Engineering, Box 352500, University of Washington, Seattle, WA 98195

We developed a personal computer–based machine vision system and applied it in computer-aided potato inspection. The system was able to classify 50 potato images per second for potato weight, cross-sectional diameter, shape, and color, which are the important criteria in sorting potatoes in practice. An ellipse was used as the shape descriptor for potato shape inspection and color thresholding was performed in the hue-saturation-value (HSV) color space to detect green color defects. Our machine vision algorithms were evaluated objectively for weight and cross-sectional diameter inspection, and subjectively for shape and color inspection on 200 potatoes of three varieties. In addition, a subjective evaluation experiment was carried out with six United States Department of Agriculture–licensed professional inspectors participating. The average success rate was 91.2% for weight inspection and 88.7% for diameter inspection. The shape and color inspection algorithms achieved 85.5% and 78.0% success rates, respectively. The overall success rate, combining all of the above criteria, was 86.5%. This type of machine vision system can be reliably used in the future to sort out the definitely good and bad potatoes and to forward the rest for human sorting, thus potentially reducing manual efforts significantly. © 1998 John Wiley & Sons, Inc. Int J Imaging Syst Technol 9: 423–433, 1998

... Zhou等^[10]开发了一个针对马铃薯质量、形状、切面直径、颜色等特征进行马铃薯分选的计算机视觉系统,分选准确率在85%以上 ...

2012

0.0

. 2012, 63(1):null-null

A real-time mathematical computer method for potato inspection using machine vision

... Navid等^[11]提出了基于机器视觉的马铃薯外部缺陷和形状的在线检测方法,形状特征选取了外接矩形长宽比、最大直径和最小直径3个特征,融合马铃薯的缺陷特征对马铃薯进行在线分选,结果证明支持向量机的方法在准确率和效率上具有较好的表现,分选准确率在95%以上 ...

2012

2.276

0.0

. 2012, 112(1-2):null-null

In-line sorting of irregular potatoes by using automated computer-based machine vision system

... Gamal等^[12]提出了一种基于机器视觉的马铃薯形状实时检测方法,该方法提取了离心率、偏心度、圆形度、伸长量4个基于傅立叶的相关参数,对马铃薯进行在线分选,总分选精度为96% ...

2010

0.0

1.299

. 2010, 26(2):347-350

Potato shape detection based on Zernike moments

目前基于机器视觉的马铃薯薯形检测的形状特征单一,相关研究较少,为了进一步探索合适的形状特征参数及检测方法,该文将Zernike矩作为特征参数并利用支持向量机实现了马铃薯薯形的检测分类,准确度较高.首先用截取最佳图像的方法对马铃薯图像进行归一化,使得归一化后的图像具有平移和尺度不变性,然后从归一化的图像中计算具有旋转不变性的Zernike矩参数,通过特征筛选确定分类的19个Zernike特征参数,最后将这些特征输入到支持向量机中,用高斯径向基核函数(RBF)和Sigmoid核函数构建混合核函数,完成马铃薯薯形检测分类,对薯形良好和畸形的检测准确率达93%和100%,能够准确剔除畸形马铃薯并满足实际检测的要求.

Abstract：

Up to now, the shape feature of potato shape detection based on machine vision is single with little relative investigation. Taking Zernike moments and support vector machine as shape detecting feature and classifier respectively, an approach to potato shape detection and classification, which yielded a relatively higher accuracy, was proposed in this paper. The image was first normalized by using best image segmentation method to obtain scale and translation invariance. The rotation invariant Zemike features were then extracted from the normalized images, among which 19 features were selected. At last, shape classification was accomplished by inputting the selected features into support vector machine classifier. A new mixed kernel function of RBF and Sigmoid kernel function was proposed, resulting in 93% and 100% detection accuracy for the perfect and malformation potatoes, respectively.

... 内蒙古农业大学郝敏等^[13]提取马铃薯的Zernike矩作为特征参数并利用支持向量机实现了马铃薯外形的分类 ...

2012

0.0

1.299

. 2012, 28(7):178-183

Automatic detecting and grading method of potatoes based on machine vision

In order to realize grading of potato, a potato grading system based on machine vision was developed. Two pieces of plane mirror placed in V-shape were used to get three surface images of a potato at one time. Volume method based on minimum circumscribed cylinder was proposed to grade potatoes according to their size, the ratio of width and length of the longest diameter circum-rectangle was used to grade potatoes according to their shape. On the basis of characteristics of potato defects, defect area, diagonal length and cross method were used as criterions of the lacunary, dry rot, mechanical damaged, budded and misshapen potatoes. Experiments showed that the recognition accuracy of potato defects was 91.0%. The results indicate that the classification method has a high accuracy, and can be used for external quality online detection of potatoes.

为了实现马铃薯的自动分级，设计了基于V型平面镜同时获取三面图像的马铃薯机器视觉分级系统，并提出了相应的分级算法。根据大小特性, 提出基于最小外接柱体体积法的马铃薯大小分级检测方法；根据外形特性，采用最长径外接矩形的宽高比法，实现了类圆形、椭圆型以及长型马铃薯的分类；根据马铃薯缺陷特点，分别提出以缺陷面积大小作为判别准则的孔洞、干腐马铃薯判别方法，以外接矩形对角线长度作为判别准则的机械损伤马铃薯判别方法和基于交叉法的发芽、畸形马铃薯检测方法。最终马铃薯的分级正确率为91.0%。试验结果表明：基于机器视觉的马铃薯自动分级检测方法可行，可用于马铃薯外部品质的在线检测。

2009

0.0

2007

0.0

... Sun等^[16]为了克服上述缺点并且降低分类算法的时间消耗,对RELIEF方法进行了改进,提出了I-RELIEF算法,本文运用了该算法对马铃薯形状特征进行加权处理 ...

0.0

... 2 支持向量机算法支持向量机是由Corinna Cortes和Vapuik在1995年首次提出来的^[17] ...