Chunjiang Zhao

Fast detection and visualization of early decay in citrus using Vis-NIR hyperspectral imaging

Early detection of fungal infection in citrus fruit is one of the major problems.Hyperspectral information visualization was first used to detect decayed citrus.Mean normalization is proposed to reduce spectral variability due to spherical fruit.Four spectral images were selected for development of multispectral algorithm.The total success rate is 98.6% for test set with no false negatives. Early detection of fungal infection in citrus fruit is one of the major problems in the postharvest phase. The automation of this task is still a challenge for the citrus industry. In this study, the potential application of hyperspectral imaging, which combines conventional imaging and spectroscopy to simultaneously acquire both spatial and spectral information from an object, was evaluated for automatic detection of the early symptoms of decay caused by Penicillium digitatum fungus in citrus fruit. Hyperspectral images of sound and decayed navel oranges were acquired in the wavelength range of 325-1100nm. Principal component analysis (PCA) was applied to a dataset comparing of average spectra from decayed and sound tissue to reduce the dimensionality of data and to observe the ability of visible-near infrared (Vis-NIR) hyper-spectra to discriminate data from two classes. And, a mean normalization step is applied prior to PCA to reduce the effect of sample curvature on spectral profiles. In this case it was observed that sound and decayed spectra were separable along the resultant first principal component (PC1) axis, then, four wavelength images centered at 575, 698, 810 and 969nm were selected as the characteristic wavelength images by analyzing the weight coefficients of PC1 in order to develop a fast classification method for establishing an on-line multispectral imaging system. Subsequently, a combination image, which obtained by multiplying the characteristic weight coefficients by corresponding to mean-normalized characteristic wavelength images of each orange sample, was calculated for determination of decayed fruits. Based on the obtained multispectral combination image, the technique of intensity slicing as one of the pseudo-color image processing methods is used to transform the combination image into a 2-D visual classification image that would enhance the contrast between sound and decayed classes. Finally, an image segmentation algorithm for detection of decayed fruit was developed based on the pseudo-color image coupled with a simple thresholding method. For the investigated 210 naval orange samples including 80 sound fruits and 130 infected fruits, the total success rate is 100% for training set and 98.6% for test set with no false negatives, respectively, indicating that the proposed multispectral algorithm here is capable of detecting decay caused by penicillium digitatum in naval orange fruit using only four key wavelength images. The results from this study could be used for development of a non-destructive monitoring system for rapid detection of decayed citrus on the processing line. The idea behind the proposed algorithm can be extended to detect the non-visible damages of other fruit, such as slight bruise and chilling injury in apples.

Hyperspectral imaging combined with multivariate analysis and band math for detection of common defects on peaches (Prunus persica)

Common defects are divided into two different types.For artificial defects, multivariate analysis were used for wavelengths selection.For non-artificial defect and stem detection, band math were constructed.The uneven lightness distribution was also investigated in this paper. Automatic detection of common defects on peaches by using imaging system is still a challenge due to the high variability of peach surface color, the similarity between the defects and stem, as well as the uneven distribution of lightness on peaches. In order to detect the common defects on peaches using hyperspectral imaging, defects were divided into two different types: artificial defects and non-artificial defects. For artificial defect detection, a two-step multivariate analysis method (Monte Carlo-Uninformative Variable Elimination and successful projections algorithm) was conducted in the spectral domain for the discriminant wavelength (DW) selection, and then minimum noise fraction (MNF) transform was conducted on the images at DWs for image processing and artificial defect detection. For the candidate non-artificial defect detection, a pair of two characteristic wavelengths at 925nm and 726nm was selected by analyzing the full spectra of sound and non-artificial defective regions, and then a band math equation was constructed for differentiating the non-artificial defect regions and stems from the sound and physical damage regions, and the candidate non-artificial defects (including non-artificial defects and stems) could be segmented by using a simple threshold method. In order to distinguish the stem from the segmented candidate non-artificial defect regions, another band math equation was constructed based on another pair of two characteristic wavelengths at 650nm and 675nm for stem identification. Additionally, the uneven lightness distribution in the spectral images was also investigated and eliminated by the band math methods. The overall classification accuracy of 93.3% for the 120 samples indicated that the selected DWs and proposed method were suitable and efficient for the common defect detection. The limitation of our research is the static inspection in one single view.

Variable Selection in Visible and Near-Infrared Spectral Analysis for Noninvasive Determination of Soluble Solids Content of ‘Ya’ Pear

Informative variable selection or wavelength selection plays an important role in the quantitative analysis of near-infrared (NIR) spectra because the modern spectroscopy instrumentations usually have a high resolution and the obtained spectral data sets may have thousands of variables and hundreds or thousands of samples. In this study, a new combination of Monte Carlo–uninformative variable elimination (MC-UVE) and successive projections algorithm (SPA; MC-UVE-SPA) was proposed to select the most effective variables. MC-UVE was firstly used to eliminate the uninformative variables in the raw spectra data. Then, SPA was applied to determine the variables with the least collinearity. A case study was done based on the NIR spectroscopy for the non-destructive determination of soluble solids content (SSC) in ‘Ya’ pear. A total of 160 samples were prepared for the calibration (n = 120) and prediction (n = 40) sets. Three calibration algorithms including linear regressions of partial least square regression (PLS) and multiple linear regression (MLR), and nonlinear regression of least-square support vector machine (LS-SVM) were used for model establishment by using the selected variables by SPA, UVE, MC-UVE, UVE-SPA, and MC-UVE-SPA, respectively. The results indicated that linear models such as PLS and MLR were more effective than nonlinear model such as LS-SVM in the prediction of SSC of ‘Ya’ pear. In terms of linear models, different variable selection methods can obtain a similar result with the RMSEP values range from 0.2437 to 0.2830. However, combination of MC-UVE and SPA was helpful for obtaining a more parsimonious and efficient model for predicting the SSC values in ‘Ya’ pear. Twenty-two effective variables selected by MC-UVE-SPA achieved the optimal linear MC-UVE-SPA-MLR model compared with other all developed models by balancing between model accuracy and model complexity. The coefficients of determination (r2), root mean square error of prediction, and residual predictive deviation by MC-UVE-SPA-MLR were 0.9271, 0.2522, and 3.7037, respectively.

Detection of bruise and stem-end/calyx of apples using hyperspectral imaging and segmented principal component analysis

Because the images of bruises often exhibit patterns and intensity values similar to the stem-end/calyx, it is important to discriminate the bruises from the stem-end/calyx in an apple sorting system. To solve this problem, a push-broom hyperspectral imaging system was developed to acquire reflectance images of apple between 400 nm and 1100 nm. A total of 60 apples were used, 30 of them were sound, 15 of them with bruises near the calyx, and 15 of them with bruises near the stem-end. The full wavelength region from 450 to 980 nm was segmented into the visible region 450 to 780 nm and the near-infrared region from 780 to 980 nm. Then the principal component analysis (PCA) was conducted on the full region and the two segmented regions respectively. The PC images were used to detect the bruise and the effective wavebands were selected according to the PC images’ loading plots. The PC images from 450 to 780 nm could not be used to detect the bruises. The PCA was used again on the effective wavebands selected from 780 to 980 nm and 450 to 980 nm. Results show that the effective wavebands from 780 to 980 nm could be used to discriminate bruises from stem-end and calyx. None of the healthy apple was misclassified. None of the bruises were misclassified as stem-end or calyx. 93.3% of the bruises near the stem-end were correctly classified, and only 86.7% of the bruises near the calyx were correctly classified using the PC images resulted from the effective wavebands 820 and 970nm. The classification error of the bruises near the stem-end/calyx would be caused by the strong light spot on the sample. Moreover, the only 2 effective wavebands 820 and 970 nm from the NIR region would decrease the cost to establish a multispectral imaging system.