基于纹理特征的网纹哈密瓜分类研究

发布时间：2018-11-20 07:01

【摘要】：我国哈密瓜产量虽然占据世界第一,但是国际市场竞争力弱,售价远低于日韩等其他国家,经济效益低下。主要因为传统网纹哈密瓜外观品质检测及分级以人工分选为主,效率低下,精度差。随着图像处理技术和计算机软硬件的日益提高,使得智能的机器视觉技术在网纹哈密瓜外部品质检测中应用越加广泛。本研究以西州蜜17号,金蜜3号和八六王三种网纹哈密瓜为研究对象,根据网纹哈密瓜重要的外部品质之一——纹理特征对哈密瓜不同品种和不同纹理等级进行分类研究,并初步探究了哈密瓜外部纹理和内部中心糖度的相关性。主要研究内容和结论有:(1)设计构建了哈密瓜图像采集装置,装置由移动托盘、彩色CCD相机、LED光源、自制光照箱和计算机等组成,可以获取完整清晰的哈密瓜图像。分析了截取ROI的必要性,并介绍了一种ROI截取方法,综合对比了 500×500、400×400、300×300、200×200和100×100像素五种不同大小的ROI截取方案,最终选择了 300×300像素的ROI图像。(2)建立了基于纹理特征的哈密瓜品种分类模型。对西州蜜17号、金蜜3号和八六王等3个品种,采用5种不同的纹理特征分析方法,提取了 84个纹理特征,发现GLCM提取的8个纹理特征能够有效区分不同品种哈密瓜的图像,预测集分类准确率达到了 98.52%。通过制定相应的映射规则,实现了哈密瓜三品种样本分类,样本分类的结果达到了 100%。实验结果说明GLCM提取的纹理特征可以较高精度的区分哈密瓜三品种的图像和样本,满足哈密瓜不同品种分类要求。(3)建立了基于纹理特征的哈密瓜纹理等级分类模型。对于特等、一等和等外3个等级纹理,对比了 SFS、GA和mRMR三种特征选择方法优选后的纹理特征对不同等级纹理的分类结果。发现采用SFS方法对组合特征的降维效果相对最好,优选后的特征数量分别为13、33和21。另外,SFS方法对三品种哈密瓜不同等级纹理图像的预测集分类准确率也相对最高,分别为90.00%、89.44%和86.67%。定义了相应的映射规则实现了哈密瓜三等级样本分类,样本的分类准确率分别为91.67%、88.33%和83.33%,与三等级纹理图像的分类结果基本接近。实验结果说明了组合纹理特征结合SFS特征优选的方法能够实现哈密瓜不同等级纹理图像和样本的分类,且对不同哈密瓜品种具有较好的鲁棒性。(4)分析了纹理特征与哈密瓜糖度的相关性。以八六王为例,从哈密瓜纹理等级和纹理特征两个方面分析与中心糖度的相关性。对比了 PLS、SMLR和PCR三种建模方法对糖度的预测结果,发现采用PLS糖度检测模型的预测结果最好,校正集和交叉验证相关系数分别为0.8804和0.7524,RMSEC为0.9476°Brix,RMSECV为1.3403°Brix。实验结果说明纹理特征与中心糖度存在一定的相关性。
[Abstract]:Although the output of Hami melon occupies the first place in the world, the competitiveness of international market is weak, the selling price is far lower than other countries, such as Japan and Korea, and the economic benefit is low. The main reason is that the traditional mesh Hami melon appearance quality detection and classification mainly by manual sorting, low efficiency and poor accuracy. With the increasing improvement of image processing technology and computer software and hardware, intelligent machine vision technology has been widely used in the external quality detection of reticulated Hami melon. In this study, three kinds of reticulated cantaloupe, Mi 17 of Xizhou, Jinmi 3 and Wang 86, were studied. According to one of the important external qualities, texture feature, the classification of different varieties and texture grades of Hami melon was carried out. The correlation between the outer texture of Hami melon and the saccharification of the inner center was also preliminarily explored. The main research contents and conclusions are as follows: (1) an image acquisition device for Hami melon is designed and constructed. The device consists of a moving tray, a color CCD camera, a LED light source, a self-made illumination box and a computer, which can obtain a complete and clear picture of Hami melon. In this paper, the necessity of intercepting ROI is analyzed, and a method of ROI interception is introduced. Five different ROI intercepting schemes of 500 脳 500400 脳 400300 脳 300200 脳 200 and 100 脳 100 pixels are compared. Finally, 300 脳 300 pixel ROI images were selected. (2) the classification model of Hami melon varieties based on texture features was established. Using 5 different texture feature analysis methods, 84 texture features were extracted from three varieties of Xizhou Mi 17, Jinmi 3 and 86 Wang. It was found that the 8 texture features extracted by GLCM could effectively distinguish the images of different varieties of Hami melon. The accuracy of prediction set classification is 98.52%. By making corresponding mapping rules, the sample classification of Hami melon was realized, and the result of sample classification reached 100. The experimental results show that the texture features extracted by GLCM can distinguish the images and samples of three varieties of Hami melon with high accuracy and meet the classification requirements of different varieties of Hami melon. (3) A texture classification model of Hami melon based on texture features is established. For the three gradation textures of principal grade, first class and equal class, the classification results of different grade textures based on SFS,GA and mRMR feature selection methods are compared. It is found that the SFS method is the best in reducing the dimension of the combined features, and the number of the selected features is 131.33 and 21 respectively. In addition, the SFS method has the highest classification accuracy for three varieties of Hami melon texture images of different grades, which is 89.44% and 86.67%, respectively. The corresponding mapping rules are defined to realize the classification of Hami melon three-level samples. The classification accuracy of the samples is 91.67 88.33% and 83.33% respectively, which is close to the classification results of the three-level texture images. The experimental results show that the combination of texture features and SFS features can achieve the classification of different texture images and samples of Hami melon. And it has good robustness to different varieties of Hami melon. (4) the correlation between texture features and sugar content of Hami melon is analyzed. Taking 86 king as an example, the correlation between central sugar content and texture grade and texture feature of Hami melon was analyzed. By comparing the prediction results of three modeling methods, PLS,SMLR and PCR, it is found that the best prediction results are obtained by using the PLS saccharification detection model. The correlation coefficients of correction set and cross validation are 0.8804 and 0.7524 respectively. The correlation coefficient of PLS is 0.9476 掳Brix,RMSECV and 1.3403 掳Brix., respectively. The experimental results show that there is a certain correlation between texture features and central saccharification.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：S652.1

【相似文献】