Citrus canker detection using hyperspectral reflectance imaging and PCA-based image classification method

Citrus canker is one of the most devastating diseases that threaten marketability of citrus crops. Technologies that can efficiently identify citrus canker would assure fruit quality and safety and enhance the competitiveness and profitability of the citrus industry. This research was aimed to investigate the potential of using hyperspectral imaging technique for detecting canker lesions on citrus fruit. A portable hyperspectral imaging system consisting of an automatic sample handling unit, a light source, and a hyperspectral imaging unit was developed for citrus canker detection. The imaging system was used to acquire reflectance images from citrus samples in the wavelength range between 400 and 900 nm. Ruby Red grapefruits with normal and various diseased skin conditions including canker, copper burn, greasy spot, wind scar, cake melanose, and specular melanose were tested. Hyperspectral reflectance images were analyzed using principal component analysis (PCA) to compress the 3-D hyperspectral image data and extract useful image features that could be used to discriminate cankerous samples from normal and other diseased samples. Image processing and classification algorithms were developed based upon the transformed images of PCA. The overall accuracy for canker detection was 92.7%. Four optimal wavelengths (553, 677, 718, and 858 nm) were identified in visible and short-wavelength near-infrared region that could be adopted by a future multispectral imaging solution for detecting citrus canker on a sorting machine. This research demonstrated that hyperspectral imaging technique could be used for discriminating citrus canker from other confounding diseases.     

Automated Systems Based on Machine Vision for Inspecting Citrus Fruits from the Field to Postharvest—a Review

Computer vision systems are becoming a scientific but also a commercial tool for food quality assessment. In the field, these systems can be used to predict yield, as well as for robotic harvesting or the early detection of potentially dangerous diseases. In postharvest handling, it is mostly used for the automated inspection of the external quality of the fruits and for sorting them into commercial categories at very high speed. More recently, the use of hyperspectral imaging is allowing the detection of not only defects in the skin of the fruits but also their association to certain diseases of particular importance. In the research works that use this technology, wavelengths that play a significant role in detecting some of these dangerous diseases are found, leading to the development of multispectral imaging systems that can be used in industry. This article reviews recent works that use colour and non-standard computer vision systems for the automated inspection of citrus. It explains the different technologies available to acquire the images and their use for the non-destructive inspection of internal and external features of these fruits. Particular attention is paid to inspection for the early detection of some dangerous diseases like citrus canker, black spot, decay or citrus Huanglongbing.     

Comparison of ROC Feature Selection Method for the Detection of Decay in Citrus Fruit Using Hyperspectral Images

Hyperspectral imaging systems allow to detect the initial stages of decay caused by fungi in citrus fruit automatically, instead of doing it manually under dangerous ultraviolet illumination, thus preventing the fungal infestation of other sound fruit and, consequently, the enormous economical losses generated. However, these systems present the disadvantage of generating a huge amount of data, which is necessary to select for achieving some result useful for the sector. There are numerous feature selection methods to reduce dimensionality of hyperspectral images. This work compares a feature selection method using the area under the receiver operating characteristic (ROC) curve with other common feature selection techniques, in order to select an optimal set of wavelengths effective in the detection of decay in a citrus fruit using hyperspectral images. This comparative study is done using images of mandarins with the pixels labelled in five different classes: two types of healthy skin, two types of decay and scars, ensuring that the ROC technique generally provides better results than the other methods.     

Application of Visible and Near-Infrared Hyperspectral Imaging for Detection of Defective Features in Loquat

The intent of present work was to develop a valid method for detection of defective features in loquat fruits based on hyperspectral imaging. A laboratorial hyperspectral imaging device covering the visible and near-infrared region of 380–1,030 nm was utilized to acquire the loquat hyperspectral images. The corresponding spectral data were extracted from the region of interests of loquat hyperspectral images. The dummy grades were assigned to the defective and normal group of loquats, separately. Competitive adaptive reweighted sampling (CARS) was conducted to elect optimal sensitive wavelengths (SWs) which carried the most important spectral information on identifying defective and normal samples. As a result, 12 SWs at 433, 469, 519, 555, 575, 619, 899, 912, 938, 945, 970, and 998 nm were selected, respectively. Then, the partial least squares discriminant analysis (PLS-DA) model was established using the selected SWs. The results demonstrated that the CARS-PLS-DA model with the discrimination accuracy of 98.51 % had a capability of classifying two groups of loquats. Based on the characteristics of image information, minimum noise fraction (MNF) rotation was implemented on the hyperspectral images at SWs. Finally, an effective approach for detecting the defective features was exploited based on the images of MNF bands with “region growing” algorithm. For all investigated loquat samples, the developed program led to an overall detection accuracy of 92.3 %. The research revealed that the hyperspectral imaging technique is a promising tool for detecting defective features in loquat, which could provide a theoretical reference and basis for designing classification system of fruits in further work.     

Development of a Hyperspectral Computer Vision System Based on Two Liquid Crystal Tuneable Filters for Fruit Inspection. Application to Detect Citrus Fruits Decay

Hyperspectral systems are characterised by offering the possibility of acquiring a large number of images at different consecutive wavebands. To ensure reliable and repeatable results using this kind of optical sensors, the intensity shown by the objects in the different spectral images must be independent from the differences in sensitivity of the system for the different wavelengths. The spectral efficiency of the acquisition devices and the spectral emission of the lighting system vary across the spectrum and the images, and therefore the results can reproduce these variations if the system is not properly calibrated and corrected. This is particularly complex, when several LCTF devices are used to obtain large spectral ranges. This work presents the development of a hyperspectral system based on two liquid crystal tuneable filters for the acquisition of images of spherical fruits. It also proposes a methodology for acquiring and segmenting images of citrus fruits aimed at detecting decay in citrus fruits that has been capable of correctly classifying 98 % of pixels as rotten or non-rotten and 95 % of fruit.     

Development of a two-band spectral imaging system for real-time citrus canker detection

Inspection of citrus canker is crucial due to its fast spread, high damage potential, and massive impact on export and domestic trade. This research was aimed to develop a prototype for real-time citrus canker detection. An inspection module was developed on a one-line commercial fruit sorting machine. Twenty tungsten halogen spotlights coupled with an aluminum dome painted with white diffuse paint provided reflectance illumination to the fruits in the detection chamber. The camera unit was a two-band spectral imaging system, which mainly consisted of a beamsplitter, two bandpass filters with central wavelengths at 730 and 830 nm, and two identical monochrome cameras. Using an exposure time of 10 ms, the imaging system can capture narrowband images without blurring from samples moving at a speed of 5 fruits/s. Spatial resolution of the acquired images was 2.3 pixels/mm. Real-time image processing and classification algorithms were developed based on a two-band ratio approach (i.e., R830/R730). The system was tested using 360 grapefruits with normal surface, canker lesions, and other peel diseases and defects. The overall classification accuracy was 95.3%, demonstrating that the methodology as well as the hardware and the software are effective and suitable for real-time citrus canker detection. Greasy spot, melanose, and sooty mold could generate false positive errors for the fruits without canker. The current system setup was limited to a single perspective view of the fruits. Future work will be conducted with an emphasis on whole surface inspection of each fruit.     

玉米籽粒霉变等级高光谱图像检测方法研究

为了快速、无损检测出储藏玉米籽粒不同霉变状况,提升玉米收储环节质检效率,尝试利用高光谱成像技术结合机器学习算法构建玉米籽粒霉变等级分类模型。采集400～1 000 nm波段范围内玉米籽粒高光谱图像,以测定的真菌孢子数为依据,将籽粒霉变状态划分为健康、轻度霉变、中度霉变和重度霉变4个等级,采用随机蛙跳(RF)算法优选出7个光谱特征变量,针对特征波段图像,利用Tamura算法共提取出21个纹理特征变量,基于颜色矩阵提取出21个颜色特征变量。进一步结合支持向量机(SVM)、极限学习机(ELM)和偏最小二乘回归(PLSR)3种算法分别建立基于光谱、图像和图谱特征融合的玉米籽粒霉变等级分类模型。经分析比较,融合光谱和图像特征并结合ELM算法建立的分类模型用于玉米籽粒霉变等级识别效果最优,训练集和测试集分类准确率(Acc)分别为94.21%和93.86%,并将玉米籽粒霉变等级进行可视化表达。     

基于深度学习的高光谱腊肉营养安全分级

本文设计的卷积神经网络-支持向量机(CNN-SVM)模型,从腊肉的高光谱成像出发,将深度学习提取特征与传统机器学习提取特征有机结合,设计出准确可靠的腊肉营养安全四分类器.利用三维卷积神经网络提取腊肉高光谱图像的深层特征,同时融合高光谱的光谱特征,联合输入支持向量机(SVM)实现对腊肉的分类和健康风险评价.结果:获得了与...     

Classification of Selected Citrus Fruits Based on Color Using Machine Vision System

This article evaluated some of the machine vision techniques to classify selected citrus fruits like oranges, sweet-lime, and lemon based on color analysis using single view fruit images. The methods carried out analyze the fruit images to extract the hue and classify using methods like color distance, linear discriminant analysis, and probability distribution function. The performance was evaluated in terms of classification accuracy relative to human classification and computational complexity. Classification accuracy above 90% could be obtained based on color classification. The probability distribution function method was found to be computationally less intensive and providing better accuracy when compared to other two methods. Since surface color is one of the indicators of maturity, color determination was also extended to find the maturity of fruits based on color variations within the same variety of fruits using red/green value ratio. The analysis also showed that the color information found in terms of hue mean and hue median can be utilized to classify the fruits based on maturity. Appropriate algorithms were developed and implemented to classify the fruits based on color and maturity using the above methods.     

Development of algorithms for detecting citrus canker based on hyperspectral reflectance imaging

BACKGROUND: Automated discrimination of fruits with canker from other fruit with normal surface and different type of peel defects has become a helpful task to enhance the competitiveness and profitability of the citrus industry. Over the last several years, hyperspectral imaging technology has received increasing attention in the agricultural products inspection field. This paper studied the feasibility of classification of citrus canker from other peel conditions including normal surface and nine peel defects by hyperspectal imaging. RESULTS: A combination algorithm based on principal component analysis and the two‐band ratio (Q_687/630) method was proposed. Since fewer wavelengths were desired in order to develop a rapid multispectral imaging system, the canker classification performance of the two‐band ratio (Q_687/630) method alone was also evaluated. The proposed combination approach and two‐band ratio method alone resulted in overall classification accuracy for training set samples and test set samples of 99.5%, 84.5% and 98.2%, 82.9%, respectively. CONCLUSION: The proposed combination approach was more efficient for classifying canker against various conditions under reflectance hyperspectral imagery. However, the two‐band ratio (Q_687/630) method alone also demonstrated effectiveness in discriminating citrus canker from normal fruit and other peel diseases except for copper burn and anthracnose. Copyright © 2011 Society of Chemical Industry     

Effect of fruit harvest time on citrus canker detection using hyperspectral reflectance imaging

The citrus industry has need for effective and efficient approaches to remove fruit with canker before they are shipped to selective international markets. The objective of this research was to study the effect of fruit harvest time on citrus canker detection using hyperspectral reflectance imaging. Ruby Red grapefruits with normal surface, canker, and five common peel diseases including greasy spot, insect damage, melanose, scab, and wind scar were collected during a 7-month harvest period. Hyperspectral reflectance images were acquired in the wavelength range of 450–930 nm. Spectral information divergence (SID) was used as a discrimination measure to perform statistical comparisons among reflectance spectra of grapefruit samples over the whole harvest season. The SID values with respect to the mean reflectance spectrum of canker for the 7-month periods were 0.0009, 0.0002, 0.0008, 0.0001, 0.0007, 0.0003, and 0.0004, respectively. Correlation analysis (CA) was used for hyperspectral band selection. Two-band ratio images using wavelengths of 729 and 834 nm selected by CA (R834/R729) gave the maximum absolute correlation value of 0.811. The mean ratio values for canker were in the range from 1.287 to 1.407, which were higher than the ratio values for other peel conditions. A simple thresholding and morphological filtering operations were applied to the two-band ratio images. The classification accuracies were in the range of 93.3–96.7% for each month. The results presented in this study demonstrated that there is no significant difference among the accuracy for canker detection over the whole harvest season using the two-band ratio images and threshold based on the spectrum of 7-month average.     

采用二次连续投影法和BP人工神经网络的寒富苹果病害高光谱图像无损检测

为提供苹果病害在线、快速、无损检测的理论依据,采用高光谱成像技术进行了北方大面积种植的寒富苹果病害无损检测研究。寒富苹果的主要病害有炭疽病、苦痘病、黑腐病和褐斑病害。为选择较少的有效波长而利于在线快速检测,首先采集高光谱苹果图像,分割出感兴趣区域并提取光谱信息,然后采用连续投影算法(successive projections algorithm,SPA)从全波长(500~970 nm)中提取了10个特征波长SPA1(502、573、589、655、681、727、867、904、942 nm和967 nm),再对这10个特征波长采用连续投影算法提取3个特征波长SPA2(681、867 nm和942 nm)。最后利用全波长光谱信息、SPA1提取的10个特征波长的光谱信息和SPA2提取的3个特征波长的光谱信息作为输入矢量采用线性判别分析、支持向量机和BP人工神经网络(BP artificial neural network,BPANN)模型进行苹果病害的检测。通过对检测结果分析,最终选择SPA2-BPANN为最佳检测方法,训练集检测率达100%,验证集检测率达100%。结果表明,高光谱成像技术可以有效对苹果病害进行检测,所获得的特征波长可为开发多光谱成像的苹果品质检测和分级系统提供参考。     

Hyperspectral Imaging-Based Classification and Wavebands Selection for Internal Defect Detection of Pickling Cucumbers

Hyperspectral imaging is useful for detecting internal defects of pickling cucumbers. The technique, however, is not yet suitable for high-speed online implementation due to the challenges in analyzing large-scale hyperspectral images. This research aimed to select the optimal wavebands from the hyperspectral image data, so that they can be deployed in either a hyperspectral or multispectral imaging-based inspection system for the automatic detection of internal defects of pickling cucumbers. Hyperspectral reflectance (400–700 nm) and transmittance (700–1,000 nm) images were acquired, using an in-house developed hyperspectral imaging system running at two conveyor speeds of 85 and 165 mm/s, for 300 “Journey” pickling cucumbers before and after internal damage was induced by mechanical load. Minimum redundancy–maximum relevance (MRMR) was used for optimal wavebands selection, and the loadings of principal component analysis (PCA) were also applied for qualitatively identifying the important wavebands that are related to the specific features. Discriminant analysis with Mahalanobis distance classifier was performed for the two-class (i.e., normal and defective) and three-class (i.e., normal, slightly defective, and severely defective) classifications using the mean spectra and textural features (energy and variance) from the regions of interest in the spectral images at selected waveband ratios. The classification results based on MRMR wavebands selection were generally better than those from PCA-based classifications. The two-band ratio of 887/837 nm from MRMR gave the best overall classification results, with the accuracy of 95.1 and 94.2 % at the conveyor speeds of 85 and 165 mm/s, respectively, for the two-class classification. The highest classification accuracies for the three-class classification based on the optimal two-band ratio of 887/837 nm were 82.8 and 81.3 % at the conveyor speeds of 85 and 165 mm/s, respectively. The mean spectra-based classification achieved better results than the textural feature-based classification, except in the three-class classification for the higher conveyor speed. The overall classification accuracies for all selected waveband ratios at the low conveyor speed were slightly higher than those at the higher conveyor speed, since the low speed resulted in more scan lines, thus higher spatial resolution hyperspectral images. The identified two-band ratio of 887/837 nm in transmittance mode could be applied for fast real-time internal defect detection of pickling cucumbers.     

基于深度学习的布匹疵点检测方法

为快速、准确检测布匹疵点,提出以深度学习目标检测框架YOLOv4为基础的布匹疵点检测方式,首先将5种常见疵点图像(吊经、百脚、结点、破洞、污渍)进行预处理,然后将图像输入到YOLOv4算法中进行分类。YOLOv4采用CSPDarknet53作为主干网络提取疵点特征,SPP模块、FPN+PAN的方式作为Neck层进行深层疵点特征提取,预测层采用3种尺度预测方式,对不同大小的疵点进行检测。研究结果表明:经600个测试集样本的验证,该方法对疵点图像的检测准确率达95%,检测单张疵点图像的速率为33 ms。与SSD、Faster R-CNN、YOLOv3方法进行比较,采用YOLOv4方法准确率更高,速度更快。     

Ripeness Classification of Astringent Persimmon Using Hyperspectral Imaging Technique

Nondestructive detection of fruit ripeness is crucial for improving fruits’ shelf life and industry production. This work illustrates the use of hyperspectral images at the wavelengths between 400 and 1,000 nm to classify the ripeness of persimmon fruit. Spectra and images of 192 samples were investigated, which were selected from four ripeness stages (unripe, mid-ripe, ripe, and over-ripe). Three classification models—linear discriminant analysis (LDA), soft independence modeling of class analogy, and least squares support vector machines were compared. The best model was LDA, of which the correct classification rate was 95.3 % with the input consisted of the spectra and texture feature of images at three feature wavelengths (518, 711, and 980 nm). Feature wavelengths selection and texture feature extraction were based on successive projection algorithm and gray level co-occurrence matrix, respectively. In addition, using the same input of ripeness detection to make an investigation on firmness prediction by partial least square analysis showed a potential for further study, with correlate coefficient of prediction set r _pre of 0.913 and root mean square error of prediction of 4.349. The results in this work indicated that there is potential in the use of hyperspectral imaging technique on non-destructive ripeness classification of persimmon. The experimental results could provide the theory support for studying online quality control of persimmon.     

基于高光谱图像光谱变量和颜色特征的霉变玉米籽粒识别

目的:准确识别霉变玉米籽粒。方法:基于高光谱图像光谱变量和颜色特征建立霉变玉米籽粒识别的新方法。先对玉米籽粒图像进行图像分割和光谱变量、颜色特征提取,并根据颜色特征生成颜色直方图;将光谱变量和颜色直方图特征组成特征集合;通过距离函数对特征集合中所有特征的分析确定霉变玉米籽粒所属类别。结果:所提方法对霉变玉米籽粒类别的最大平均识别偏差为1.12,最佳平均识别准确率为97.59%;与基于高光谱图像+随机蛙跳+极限学习机的方法、基于高光谱图像+稀疏自动编码器+卷积神经网络的方法、基于高光谱图像+蚁群优化+BP神经网络的方法相比,研究所提方法对霉变玉米籽粒类别的识别准确率明显提高。结论:该方法可实现被测玉米籽粒样品是否霉变以及霉变程度的准确判断。     

应用积分图的织物瑕疵检测快速算法

为解决现有基于图像处理的织物瑕疵检测算法实时性较差、正确率偏低等问题,提出一种包含学习和检测2个阶段的瑕疵检测算法。通过对无瑕疵模板图像的梯度能量特征及其分布特性的学习,自适应获得检测阶段所需的参数。一方面利用积分图原理将任意大小的图像块内的求和运算化简为三次加法运算,快速提取织物图像的梯度能量特征,实现织物瑕疵的实时检测,另一方面利用核函数拟合特征参数分布,结合均值漂移法求解分布峰值获得自适应的瑕疵判定阈值参数,实现织物瑕疵的准确分割。通过实验将本文算法与现有基于局部二值模式特征、小波特征、规则带特征等算法进行对比,针对包含3种纹理6类瑕疵的织物图像数据集的测试结果显示,本文算法平均处理时间为56ms,正确率为97%。     

基于高光谱成像技术的灵武长枣常见缺陷检测

以灵武长枣为研究对象,利用高光谱成像技术结合主成分分析法(principal component analysis,PCA)和最小噪声分离法(minimum noise fraction,MNF)对长枣缺陷进行快速检测与识别,主要探讨样本背景对缺陷识别的影响。首先,采集虫眼、裂痕、正常枣的高光谱图像,利用PCA法和MNF法分别对其降维去噪,选择虫眼与正常枣的PC1和M1图像、裂痕枣的PC2和M2图像进行缺陷识别,经PCA分析后的缺陷识别率均为100%,MNF处理后的识别率分别为69.2%,56.8%,100%;随后对其高光谱图像进行掩模去背景,再对其降维去噪后检测识别,PCA后的识别率均为100%,MNF后的识别率分别为73.1%,65.9%,100%。结果表明:利用高光谱成像技术结合两种降维去噪法对长枣常见缺陷的识别是可行的,背景干扰对于PCA法的缺陷识别不影响,其识别效果优于MNF法,且去背景后的MNF法缺陷识别率较未去背景的有所提高,为后续长枣缺陷的在线检测提供理论依据。     

高光谱亚像元分解预测花生中的黄曲霉毒素B1

黄曲霉毒素B1是一种剧毒、强致癌物质,具有紫外荧光特性。为研究高光谱成像技术对黄曲霉素的检测能力,在365 nm紫外灯下,通过高光谱成像系统采集5个浓度共250个花生籽粒样本33个波段(400~720 nm)的高光谱图像。提出一种基于高光谱亚像元分解丰度图像直方图量化特征预测黄曲霉毒素含量的方法。该方法首先通过N-FINDR端元提取方法获得黄曲霉毒素端元光谱,然后对高光谱图像进行非负矩阵分解(NMF),得到黄曲霉毒素丰度图像,对丰度图像构建直方图量化特征,使用偏最小二乘回归(PLS)和支持向量机回归(SVR)进行黄曲霉毒素丰度反演,五折交叉验法得到平均两种回归模型预测相对误差分别为29.95%和12.16%,RMSE最高为0.0306。本研究结果对农产品籽粒黄曲霉毒素光学快速检测具有积极意义。     

高光谱成像技术检测油茶果成熟度

为解决油茶果采摘期判断不准确可能导致的茶油产量降低问题,应用高光谱成像技术结合化学计量法对油茶果成熟度进行定性判别。完成了高光谱图像的曲率校正,分析不同成熟阶段油茶果的光谱特征和理化特征的变化情况。使用4 种不同的分类算法建立基于全波段光谱数据的油茶果成熟度判别模型,发现支持向量机（support vector machine,SVM）模型的分类正确率最高为97%。结合5 种特征变量选择方法对全波段光谱数据进行降维,发现经过竞争性自适应重加权算法（competitive adaptive reweighted sampling,CARS）选择的特征波长建立的模型正确率最高为82%。提取高光谱图像中的颜色特征和纹理特征建立SVM模型后发现,融合颜色特征和光谱特征建立的SVM模型的正确率高于使用单一的光谱特征（经CARS降维）建立的模型正确率：训练集分类正确率为95%,测试集正确率为93%。结果表明,利用高光谱成像技术能够对不同成熟度的油茶果进行较准确的分类,为茶农对油茶果最佳采摘期的判断提供科学依据,在保障茶籽产量最大化、油质最优化等方面具有重要意义。