鲜杏表面缺陷的高光谱成像检测研究
发布时间：2018-10-24 06:56
【摘要】：杏是新疆特色水果之一,其营养成分高又可作医药用途,备受消费者青睐。鲜杏分级是产后处理的核心环节,其中,鲜杏表面缺陷是等级划分的重要评定指标之一。目前国内鲜杏和杏制品加工环节缺少成熟的自动化检测分级技术,主要依靠人工进行鲜杏缺陷的分选,效率低,劳动强度大且品质难以掌控。由于高光谱成像技术将图像和光谱融合为一的特点,因此,本文利用该技术对鲜杏的表面缺陷进行快速的无损检测探究。本文选取新疆鲜杏作为研究对象,基于可见/近红外高光谱成像技术从图像和光谱两个角度对正常鲜杏、磨伤鲜杏、霉变鲜杏和虫伤鲜杏进行检测研究,探求一种快速有效识别正常鲜杏和缺陷鲜杏的无损检测方法,为搭建适于鲜杏缺陷的多光谱在线检测系统奠定理论基础。本文主要方法及研究结果如下:(1)确定了采集鲜杏图像的具体参数,采集正常、磨伤、霉变和虫伤四种不同类型鲜杏的高光谱图像,对鲜杏图像的感兴趣区域(ROI)进行手动提取,得到了各个波段下所对应的光谱数据。(2)对不同类型鲜杏的原始光谱数据,进行S-G卷积平滑、标准正态变量(SNV)和多元散射校正(MSC)3种不同的预处理。对比了三种预处理方法对于SVM建模的影响。结果表明,原始光谱和S-G卷积平滑预处理建立的C-SVC类型的支持向量机模型较优,识别率均为93.3%。(3)对全波段鲜杏光谱数据采用PCA降维,通过权重系数优选出495nm、570nm、729nm和891nm四个特征波段。分别对比了全波段和特征波段采用支持向量机、偏最小二乘判别、BP神经网络三种不同的判别方法的分类识别效果。结果表明:基于特征波段构建的SVM预测模型,其识别结果整体上优于全波段构建的SVM预测模型,其中,对优选的特征波段采用线性核函数构建的C-SVC类型SVM的识别率为100%;全波段构建的PLS-DA的检测判别结果优于采用特征波段构建的检测判别结果;利用全波段构建的BP神经网络识别效果优于特征波段构建的BP神经网络识别效果。另外,利用特征波段构建的SVM的识别效果优于BP神经网络和PLS-DA。(4)对四种不同类型鲜杏高光谱图像进行全波段主成分分析和特征波段主成分分析,选择缺陷部位明显的PC图像检测判别。结果表明:利用全波段主成分分析,正常鲜杏和霉变鲜杏的分类准确率较好,均达到100%,虫伤鲜杏的判别率为88.3%,磨伤鲜杏识别率最低,仅为38.3%;采用特征波段主成分分析进行检测识别,正常鲜杏、霉变鲜杏和虫伤鲜杏识别率分别为100%、100%、95%,磨伤鲜杏的判别准确率提高到80%;全波段主成分分析的整体识别率为81.7%,而特征波段主成分分析整体识别率提高到93.8%。(5)对四种不同类型鲜杏高光谱图像基于全波段和特征波段进行最小噪声分离变换,选择缺陷部位明显的MNF图像进行缺陷识别。结果表明:基于全波段的MNF,四种类型鲜杏的识别效果均较低,正常鲜杏和磨伤鲜杏的识别率分别为38.3%和33.3%,霉变鲜杏和虫伤鲜杏识别率分别为53.3%、50%;基于特征波段的MNF,正常鲜杏和虫伤鲜杏识别率提高到73.3%和71.7%;全波段最小噪声分离处理的整体识别率为43.8%,而特征波段最小噪声分离处理整体识别率为60%。通过比较,主成分分析的整体识别效果优于最小噪声分离处理的整体识别效果;对优选的特征波段进行PCA和MNF,其整体识别率分别提高到93.8%和60%;由此说明,特征波段的PCA可较有效地识别缺陷鲜杏和正常鲜杏。(6)为进一步提高鲜杏缺陷的检测率,尝试对二次主成分分析检测后未识别的磨伤鲜杏进行图像的波段比运算。对4个特征波段下对应的图像进行两两组合,选择570nm/891nm波段比图像进行检测。结果表明:磨伤鲜杏识别率由80%提高到88.3%。
[Abstract]:Apricot is one of Xinjiang's characteristic fruits, its nutrient content is high and can be used for medical use, which is favored by consumers. The classification of fresh apricot is the key link of post-treatment, among which the surface defect of fresh apricot is one of the important indexes of classification. At present, the processing links of fresh apricot and apricot products lack mature automatic detection and classification technology, and mainly rely on the sorting of fresh apricot defect artificially, the efficiency is low, the labor intensity is large, and the quality is difficult to control. Because hyperspectral imaging technology combines the image and the spectrum into one feature, this paper makes use of this technique to detect the surface defects of fresh apricot. In this paper, Xinjiang fresh apricot is selected as the research object, and based on the visible/ near-infrared hyperspectral imaging technology, the normal fresh apricot, the fresh apricot, the mildewed fresh apricot and the insect-wound fresh apricot are detected by the visible/ near-infrared hyperspectral imaging technology, A nondestructive testing method for quickly and effectively identifying normal fresh apricot and defective fresh apricot is studied, which lays a theoretical foundation for constructing multi-spectrum on-line detection system suitable for fresh apricot defect. The main methods and research results are as follows: (1) the specific parameters of the fresh apricot image are determined, the high-spectrum images of four different types of fresh apricot are collected, the region of interest (ROI) of the fresh apricot image is manually extracted, the corresponding spectral data in each band is obtained. (2) The original spectral data of different kinds of fresh apricot were processed by S-G convolution smoothing, standard positive-state variable (SNV) and multi-scatter correction (MSC). The effects of three pretreatment methods on SVM modeling were compared. The results show that the support vector machine model of C-SVC type established by the original spectrum and S-G convolution smoothing pre-processing is better, and the recognition rate is 93. 3%. (3) Using PCA to reduce the spectral data of the whole band, the four characteristic bands of 495nm, 570nm, 729nm and 891nm are determined by weight coefficient. The classification and recognition effects of three different discrimination methods of support vector machine, partial least two multiplication discrimination and BP neural network are compared respectively in the full band and the characteristic band. The results show that SVM prediction model based on feature band construction is superior to SVM prediction model constructed by full band, and the recognition rate of C-SVC type SVM constructed by linear kernel function for the preferred feature band is 100%. The detection result of PLS-DA constructed by the whole band is superior to the detection discrimination result constructed by the characteristic wave band, and the BP neural network identification effect constructed by the whole band is superior to the BP neural network identification effect constructed by the characteristic band. In addition, the recognition effect of SVM based on feature band is better than that of BP neural network and PLS-DA. (4) carrying out full band main component analysis and characteristic band main component analysis on four different types of fresh apricot high-spectrum images, and selecting obvious PC image detection discrimination on the defect parts. The results showed that the classification accuracy of fresh apricot and moldy fresh apricot reached 100% with the analysis of the main components of the whole band. The discrimination rate of fresh apricot was 88. 3%, the recognition rate of fresh apricot was 38. 3%, and the main component analysis of the characteristic bands was used to detect and identify the fresh apricot. The recognition rate of fresh apricot and worm was 100%, 100%, 95% respectively, and the accuracy of discrimination of fresh apricot was improved to 80%. The overall recognition rate of the whole band principal component analysis was 81.7%, while the overall recognition rate of the characteristic band principal component analysis was improved to 93.8%. and (5) carrying out minimum noise separation conversion on the four different types of fresh apricot high-spectrum images on the basis of the whole band and the characteristic wave band, and selecting the MNF images with obvious defect parts to perform defect identification. The results showed that the recognition rate of fresh apricot and fresh apricot was 35.3% and 33.3% respectively, and the recognition rate of fresh apricot and fresh apricot was 53.3% and 50%, respectively. The recognition rate of fresh apricot and fresh apricot was improved to 73. 3% and 71.7%, the overall recognition rate of the whole band minimum noise separation processing was 43.8%, and the overall recognition rate of the characteristic band minimum noise separation processing was 60%. By comparison, the overall recognition effect of the principal component analysis is better than the overall recognition effect of the minimum noise separation process; PCA and MNF are performed on the preferred characteristic bands, and the overall recognition rate is improved to 93.8% and 60%, respectively; therefore, The PCA of the characteristic wave band can effectively identify the defective fresh apricot and the normal fresh apricot. and (6) in order to further improve the detection rate of the fresh apricot defect, try to carry out the band ratio calculation of the image of the fresh apricot which is not recognized after the secondary main component analysis and detection. two combinations of the corresponding images in four feature bands are performed, and the 570nm/ 891nm wave band is selected to be detected than the image. The results showed that the identification rate of fresh apricot was increased from 80% to 88. 3%.
【学位授予单位】：石河子大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;S662.2
[Abstract]:Apricot is one of Xinjiang's characteristic fruits, its nutrient content is high and can be used for medical use, which is favored by consumers. The classification of fresh apricot is the key link of post-treatment, among which the surface defect of fresh apricot is one of the important indexes of classification. At present, the processing links of fresh apricot and apricot products lack mature automatic detection and classification technology, and mainly rely on the sorting of fresh apricot defect artificially, the efficiency is low, the labor intensity is large, and the quality is difficult to control. Because hyperspectral imaging technology combines the image and the spectrum into one feature, this paper makes use of this technique to detect the surface defects of fresh apricot. In this paper, Xinjiang fresh apricot is selected as the research object, and based on the visible/ near-infrared hyperspectral imaging technology, the normal fresh apricot, the fresh apricot, the mildewed fresh apricot and the insect-wound fresh apricot are detected by the visible/ near-infrared hyperspectral imaging technology, A nondestructive testing method for quickly and effectively identifying normal fresh apricot and defective fresh apricot is studied, which lays a theoretical foundation for constructing multi-spectrum on-line detection system suitable for fresh apricot defect. The main methods and research results are as follows: (1) the specific parameters of the fresh apricot image are determined, the high-spectrum images of four different types of fresh apricot are collected, the region of interest (ROI) of the fresh apricot image is manually extracted, the corresponding spectral data in each band is obtained. (2) The original spectral data of different kinds of fresh apricot were processed by S-G convolution smoothing, standard positive-state variable (SNV) and multi-scatter correction (MSC). The effects of three pretreatment methods on SVM modeling were compared. The results show that the support vector machine model of C-SVC type established by the original spectrum and S-G convolution smoothing pre-processing is better, and the recognition rate is 93. 3%. (3) Using PCA to reduce the spectral data of the whole band, the four characteristic bands of 495nm, 570nm, 729nm and 891nm are determined by weight coefficient. The classification and recognition effects of three different discrimination methods of support vector machine, partial least two multiplication discrimination and BP neural network are compared respectively in the full band and the characteristic band. The results show that SVM prediction model based on feature band construction is superior to SVM prediction model constructed by full band, and the recognition rate of C-SVC type SVM constructed by linear kernel function for the preferred feature band is 100%. The detection result of PLS-DA constructed by the whole band is superior to the detection discrimination result constructed by the characteristic wave band, and the BP neural network identification effect constructed by the whole band is superior to the BP neural network identification effect constructed by the characteristic band. In addition, the recognition effect of SVM based on feature band is better than that of BP neural network and PLS-DA. (4) carrying out full band main component analysis and characteristic band main component analysis on four different types of fresh apricot high-spectrum images, and selecting obvious PC image detection discrimination on the defect parts. The results showed that the classification accuracy of fresh apricot and moldy fresh apricot reached 100% with the analysis of the main components of the whole band. The discrimination rate of fresh apricot was 88. 3%, the recognition rate of fresh apricot was 38. 3%, and the main component analysis of the characteristic bands was used to detect and identify the fresh apricot. The recognition rate of fresh apricot and worm was 100%, 100%, 95% respectively, and the accuracy of discrimination of fresh apricot was improved to 80%. The overall recognition rate of the whole band principal component analysis was 81.7%, while the overall recognition rate of the characteristic band principal component analysis was improved to 93.8%. and (5) carrying out minimum noise separation conversion on the four different types of fresh apricot high-spectrum images on the basis of the whole band and the characteristic wave band, and selecting the MNF images with obvious defect parts to perform defect identification. The results showed that the recognition rate of fresh apricot and fresh apricot was 35.3% and 33.3% respectively, and the recognition rate of fresh apricot and fresh apricot was 53.3% and 50%, respectively. The recognition rate of fresh apricot and fresh apricot was improved to 73. 3% and 71.7%, the overall recognition rate of the whole band minimum noise separation processing was 43.8%, and the overall recognition rate of the characteristic band minimum noise separation processing was 60%. By comparison, the overall recognition effect of the principal component analysis is better than the overall recognition effect of the minimum noise separation process; PCA and MNF are performed on the preferred characteristic bands, and the overall recognition rate is improved to 93.8% and 60%, respectively; therefore, The PCA of the characteristic wave band can effectively identify the defective fresh apricot and the normal fresh apricot. and (6) in order to further improve the detection rate of the fresh apricot defect, try to carry out the band ratio calculation of the image of the fresh apricot which is not recognized after the secondary main component analysis and detection. two combinations of the corresponding images in four feature bands are performed, and the 570nm/ 891nm wave band is selected to be detected than the image. The results showed that the identification rate of fresh apricot was increased from 80% to 88. 3%.
【学位授予单位】：石河子大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;S662.2
【参考文献】
相关期刊论文 前10条
1 高琦煜;张赛;;基于中值滤波与维纳复原的图像预处理方法[J];电脑知识与技术;2016年34期
2 康自虎;;MATLAB在化学工程与工艺实验数据处理中的应用[J];当代化工研究;2016年09期
3 李火青;刘永强;;基于ENVI/IDL的植被指数反演系统设计[J];实验室研究与探索;2016年07期
4 白t，
本文编号：2290645
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2290645.html
