【作者】 郝敏；

【导师】 麻硕士；

【作者基本信息】 内蒙古农业大学 ， 农业电气化与自动化， 2009， 博士

【摘要】 马铃薯的品质关系到商品的经济效益和加工的复杂性,马铃薯的检测与分级一直是商品化处理中必要的处理步骤。鉴于目前国内尚无基于机器视觉的马铃薯品质检测的成熟技术,而人工检测的效率低、主观性强,本文构建了马铃薯外部品质检测的机器视觉硬件系统,对马铃薯的外部品质检测方法进行了系统研究,实现了基于机器视觉技术的重量、薯形、外部缺陷三个指标的马铃薯外观品质检测。1.设计了基于机器视觉的静态马铃薯外部品质检测硬件系统,确定了用于马铃薯图像采集的背景颜色和光照条件;依据张正友的平面模板标定法进行摄像机标定,结果表明摄相机参数可以满足检测要求。2.在重量检测中,根据马铃薯个体差异很大的特点,本文将俯视、侧视图像相结合,以俯视面积和侧面厚度两个参数建立重量检测模型,有效减少单薯误差,可实现性好,模型回归相关系数为0.9836,对大、中、小3种规格马铃薯的检测准确率分别达到了97%、96%和98%。3.在形状检测中,选择Zernike矩作为形状特征,在初始形状特征计算、特征提取以及薯形检测三个方面进行了重点研究。①提出了Zernike矩的快速计算方法,将径向多项式的系数和半径的幂次以矩阵形式存储,根据Zernike矩的参数n和m查询计算,减少了系数计算和半径幂次重复计算,对同一幅图像的多阶Zernike矩计算速度明显提高。在图像归一化中提出截取最佳图像的方法使目标位于r≤0.9的圆形区域中且目标最大化,避免了尺度归一化中形心位置的不准确和插值的影响,并用改进的Zernike矩快速计算方法和截取最佳图像的归一化方法计算初始形状特征。②研究了与遗传算法相结合用于特征选择的分类器选用原则,以分类正确率作为适应度函数的主要参数,将遗传算法与概率神经网络结合,形成一种新的特征选择方法。综合考虑分类正确率、所选择的特征个数及二者对函数影响的大小三个因素,提出了遗传算法与分类器相结合进行特征选择算法中适应度函数的设计准则,并设计了一个新的适应度函数。用改进的特征选择方法从47个初始特征中提取19个Zernike矩参数作为形状分类特征。③用支持向量机作分类器,用RBF核函数和Sigmoid核函数构建了一个新的混合核函数进行形状检测,对马铃薯形状良好和畸形的检测准确率分别为93%和100%。4.在外部缺陷检测中,分别用颜色分量法和SUSAN算子进行马铃薯外部缺陷的分割,并通过设置面积阈值实现缺陷的初步判定。更多还原

【Abstract】 The potato’s quality is directly related to the product economic benefits and processing complexity. The quality detection and classification is an essential processing step in the process of commercialization. There has been little internal research on the quality detection of potato based on machine vision. Owing to the low accuracy and strong subjectivity, a systematic study is made on the potato external quality based on the machine vision technology, to propose a set of detection criterions for evaluating the potato’s external quality comprehensively, including weight, shape and external defects.1. A hardware system based on machine vision is built for identifying potatoes’external quality. The most appropriate light source and background color are determined by a lot of experiments and comparisons. According to the plane calibration of Zhang algorithm, camera calibration is done for correcting image distortion, and its accuracy is able to satisfy the demand of identifying external quality of potatoes.2. Through combining the top view and side elevation,the projection area and side elevation thickness of potato image are extracted to classify potato weight instead of metage. The correlation coefficient of model regression is 0.9836. The detection accuracies of 98%, 96% and 97% can be gained for bigger, normal and smaller potatoes, perspectively.3. In shape detection,Zernike moments are taken as the shape feature.And the study was emphasized on three content,including the calculation of Zernike moments,features selection and shape classification.(1) We propose fast calculation method of radial polynomial in the Zernike moments, which stores the power of polynomials coefficient and radius in the form of matrix. According to parameters n & m of Zernike moments to query calculation, the coefficients calculation and repetitious calculation of the radius power is reduced with a result of improving the calculation speed of Zernike moments obviously. A new image normalization method based on optimum image cut was proposed,this method is to make the object locate in the circle range of r≤0.9 and to be the maximum.The experimental results show that this way can near-perfectly preserve the invariance of scale and rotation for potato image. The fast algorithm for computation of Zernike moments and normalization method are used to calculate the initial shape feature.(2)The classifier selecting principle is investigated by combining with the genetic algorithm. A new feature selection method with the combination of genetic algorithms and probabilistic neural networks is put forward firstly. Comprehensively considering the factor of classification accuracy,selected feature number and the impact of two factors, a new fitness function is proposed. The simulation tests indicate that the fitness function and feature selection method can be used for searching the best feature combination.The initial Zernike moments parameters of potatoes are optimized using improved genetic algorithm, and nineteen Zernike moments are extracted to form the shape feature.(3) Taking SVM as classifier, a new mixed kernel function of RBF and Sigmoid kernel function is proposed, resulting in 93% and 100% detection accuracy, respectively. Its detection accuracy can reach respectively for the perfect and malformation potatoes.4. The external defects of potatoes are segmented by using color component method and SUSAN operator respectively. Meanwhile, the defects decision is realized through setting area threshold.更多还原