Detection of insect-damaged wheat kernels using near-infrared hyperspectral imaging

Insect damage in wheat adversely affects its quality and is considered one of the most important degrading factors in Canada. The potential of near-infrared (NIR) hyperspectral imaging for the detection of insect-damaged wheat kernels was investigated. Healthy wheat kernels and wheat kernels visibly damaged by Sitophilus oryzae, Rhyzopertha dominica, Cryptolestes ferrugineus, and Tribolium castaneum were scanned in the 1000–1600 nm wavelength range using an NIR hyperspectral imaging system. Dimensionality of the acquired hyperspectral data was reduced using multivariate image analysis. Six statistical image features (maximum, minimum, mean, median, standard deviation, and variance) and 10 histogram features were extracted from images at 1101.69 and 1305.05 nm and given as input to statistical discriminant classifiers (linear, quadratic, and Mahalanobis) for classification. Linear discriminant analysis and quadratic discriminant analysis classifiers correctly classified 85–100% healthy and insect-damaged wheat kernels.     

Categorization of pork quality using Gabor filter-based hyperspectral imaging technology

Objective assessment of pork quality is important for meat industry application. Previous studies using spectral approaches focused on using color and exudation features for the determination of pork quality levels without considering the image texture feature. In this study, a Gabor filter-based hyperspectral imaging technique was presented to develop an accurate system for pork quality level classification. Texture features were obtained by filtering hyperspectral images with two-dimensional Gabor functions. Different spectral features were extracted from Gabor-filtered images and hyperspectral images. The principal component analysis (PCA) was used to compress spectral features over the entire wavelengths (400–1000 nm) into principal components (PCs). ‘Hybrid’ PCs were created by combining PCs from hyperspectral images with PC(s) from Gabor-filtered images. Both K-means clustering and linear discriminant analysis (LDA) were applied to classify pork samples. Results showed that, the accuracy of K-mean clustering analysis reached 78% by 5 hybrid PCs and 83% by 10 hybrid PCs, which were 15% and 28% higher than the results without using texture features. The highest classification accuracy using LDA reached 100% by 5 hybrid PCs. Furthermore, the cross-validation technique was applied for evaluating how the classification results would generalize to independent pork sample sets. A total of 210 partitions (different training and testing sets) were used to obtain the unbiased statistical classification results. The overall classification accuracy reached 84 ± 1% (mean ± 95% confidence interval) by 5 hybrid PCs and was 72 ± 2% when no Gabor filter-based spectral features were used. Thus, a statistically significant improvement was achieved using image texture features. Moreover, the classification results strongly suggested that the texture features along the direction of π/4 offered more useful information for the differentiation of the four main levels of pork quality.     

Authenticity green coffee bean species and geographical origin using near-infrared spectroscopy combined with chemometrics

To prevent the adulteration of agricultural resources and provide a solution to enhance the green coffee bean supply chain, authentication using the near-infrared spectroscopy (NIRS) technique was investigated. Partial least square with discrimination analysis (PLS-DA) models combined with various preprocessing methods were built from NIR spectra of 153 Vietnamese green coffee samples. The model combined with the standard normal variate and the first order of derivative yielded excellent performance in predicting coffee species with the error cross-validation of 0.0261. PLS-DA model of mean centre and first-order derivative spectra also yielded good performance in verifying geographical indication of green coffee with the error of 0.0656. By contrast, the predicting abilities of post-harvest methods were poor. The overall results showed a high potential of the NIRS in online authentication practices.     

近红外光谱结合化学计量学方法在油脂检测中的应用

近红外光谱结合化学计量学技术由其独特的技术优势在油脂掺假分析领域已取得广泛应用。本文主要对近红外光谱结合化学计量学技术的原理及特点进行论述,并综述了近年来在油脂定性及定量分析方法中应用较为广泛、实用性较强的化学计量学方法,即聚类分析、主成分分析、支持向量机、人工神经网络、主成分回归以及偏最小二乘法结合近红外光谱在油脂掺假分析中的应用,并展望了该技术的发展前景。     

近红外光谱结合化学计量学技术在食品掺假分析中的应用

近红外光谱结合化学计量学技术由其独特的技术优势在食品掺假分析领域已得到广泛应用。本文对近红外光谱结合化学计量学技术的原理及特点进行论述, 并综述了近10年来应用较为广泛、实用性较强的化学计量学方法, 偏最小二乘法、人工神经网络及支持向量机结合近红外光谱技术在食品掺假分析中应用, 展望了该技术的发展前景。     

Detection of fungal infection in five different pulses using near-infrared hyperspectral imaging

The five major pulse crops grown in Canada are: chick peas, green peas, lentils, pinto beans and kidney beans. Potential causes of fungal infection in these pulses are Aspergillus flavus and Penicillium commune. Early stages of fungal infections in pulses are not detectable with human eyes and traditional microbial methods require significant time to detect fungal infection. Near-infrared (NIR) hyperspectral imaging system is an advanced technique widely being assessed for detection of insect infestation and fungal infection in cereal grains and oilseeds. The primary objective of this study was to assess the feasibility of the NIR hyperspectral imaging system to identify fungal infections in pulses. Hyperspectral images of healthy and fungal infected chick peas, green peas, lentils, pinto beans and kidney beans were acquired and features (six statistical and 10 histogram) were used to develop classification models to identify fungal infection caused by A. flavus and P. commune. Images of healthy and fungal-infected kernels were acquired at 2 week intervals (0, 2, 4, 6, 8 and 10 weeks from artificial inoculation). Six-way (healthy vs the five different stages of infection) and two-way (healthy vs every stage of infection) models were developed and classifications were done using linear discriminant analysis (LDA) and quadratic discriminant analysis (QDA) classifiers. The LDA classifier identified both types of fungal infections with 90–94% accuracy while using the six-way model, and with 98–100% accuracy when using the two-way models for all five types of pulses. The QDA classifier also showed promising results as it gave 85–90% accuracy for the six-way model and 96–100% accuracy for the two-way models. The two fungal species could not be differentiated by the hyperspectral imaging.     

Detection of fusarium damage in Canadian wheat using visible/near-infrared hyperspectral imaging

Fusarium damage in wheat may reduce the quality and safety of food and feed products. In this study, the use of hyperspectral imaging was investigated to detect fusarium damaged kernels (FDK) in Canadian wheat samples. More than 5,200 kernels, representing seven major Canadian wheat classes, with varying degree of infection symptoms ranging from sound through mild to severe were imaged in the visible-NIR (400–1,000 nm) wavelength range. Partial least squares discriminant analysis (PLS-DA) was used to segregate kernels into sound and damaged categories based on kernel mean spectra. A universal PLS-DA model based on four wavelengths was able to detect FDK in all seven classes with an overall accuracy of 90 % and false positives of 9 %.     

Detection of different stages of fungal infection in stored canola using near-infrared hyperspectral imaging

Near-infrared (NIR) hyperspectral imaging system was used to detect different stages of fungal infections in stored canola. Artificially infected canola seeds (Fungi: Aspergillus glaucus and Penicillium spp) were subjected to hyperspectral imaging in the range between 1000 and 1600 nm at 61 evenly distributed wavelengths. Four wavelengths 1100, 1130, 1250 and 1300 nm were identified as significant wavelengths and were used in statistical discriminant analysis. Pair-wise, two-class and six-class classification models were developed to classify the healthy and different stages of fungal infected samples. Linear, quadratic and Mahalanobis discriminant classifiers were used to classify healthy, five stages of A. glaucus and five stages of Penicillium spp infected canola seeds. All the three classifiers classified healthy and fungal infected canola seeds with a classification accuracy of more than 95% for healthy canola seeds and more than 90% for the initial stages of A. glaucus and Penicillium spp infected canola seeds. The classification accuracy increased to 100% with increase in fungal infection level (length of time since inoculation). All the samples subjected to imaging were tested for seed germination and free fatty acid value (FAV). The germination decreased with increase in amount of fungal infection, whereas FAV increased with increase in amount of fungal infection.     

Detection of fungal infection and Ochratoxin A contamination in stored wheat using near-infrared hyperspectral imaging

A study was done to detect Aspergillus glaucus, and Penicillium spp., infection and Ochratoxin A contamination in stored wheat using a Near-Infrared (NIR) Hyperspectral Imaging system. Fungal-infected samples were imaged every two weeks, and the three dimensional hypercubes obtained from image data were transformed into two dimensional data. Principal component analysis was applied to the two dimensional data and based on the highest factor loadings, 1280, 1300, and 1350 nm were identified as significant wavelengths. Six statistical features and ten histogram features corresponding to the significant wavelengths were extracted and subjected to linear, quadratic and Mahalanobis discriminant classifiers. All the three classifiers differentiated healthy kernels from fungal-infected kernels with a classification accuracy of more than 90%. The quadratic discriminant classifier provided classification accuracy higher than the linear and Mahalanobis classifiers for pair-wise, two-way and six-way classification models. The Ochratoxin A contaminated samples had a unique significant wavelength at 1480 nm in addition to the two significant wavelengths corresponding to fungal infection. The peak at 1480 nm was identified only in the Ochratoxin A contaminated samples. The Ochratoxin A contaminated samples can be detected with 100% classification accuracy using NIR hyperspectral imaging system. The NIR hyperspectral system can differentiate between different fungal infection stages and different levels of Ochratoxin A contamination in stored wheat.     

Moisture content detection of maize seed based on visible/near-infrared and near-infrared hyperspectral imaging technology

To realise accurate and nondestructive detection on moisture content of maize seed based on visible/near-infrared (Vis/NIR) and near-infrared (NIR) hyperspectral imaging technology, the hyperspectral images on two sides (embryo and endosperm sides) of each maize seed of four varieties were collected. The effects of average spectra extraction regions, that is centroid region and whole seed region, and different spectral preprocessing methods, were investigated. Uninformative variable elimination (UVE) was used to extract the feature wavelengths, and the partial least squares regression (PLSR) prediction models were established. The results showed that extracting the average spectra from the centroid region did better than from the whole seed region, and S-G smoothing was prior to other preprocessing methods. The PLSR models established with NIR spectra had better performance than that with Vis/NIR spectra. The model developed for a single variety was superior to that for all varieties together.     

Rapid and real-time prediction of lactic acid bacteria (LAB) in farmed salmon flesh using near-infrared (NIR) hyperspectral imaging combined with chemometric analysis

In this study, hyperspectral imaging working in near-infrared (NIR) region (900–1700 nm) was applied to evaluate surface lactic acid bacteria (LAB) spoilage of farmed salmon flesh during cold storage. Hyperspectral images of salmon samples were acquired at different storage times. Spectral information within regions of interest (ROIs) of images were extracted to relate to reference LAB values measured by standard pour plate method. Least-squares support vector machine (LS-SVM) algorithm was used to calibrate the full NIR range spectral data, resulting in regression coefficients of prediction (R_P) of 0.929 with root mean square error of prediction (RMSEP) of 0.515. Competitive adaptive reweighted sampling (CARS) algorithm was employed to reduce the spectral redundancy and identify the most informative wavelengths (MIWs) most related with LAB prediction across the whole wavelength range. Eight individual MIWs at 1155 nm, 1255 nm, 1373 nm, 1376 nm, 1436 nm, 1641 nm, 1665 nm and 1689 nm were finally selected from the full 239 wavelengths. Based on the selected MIWs, a new optimised model named CARS-LS-SVM was established, leading to R_P of 0.925 with RMSEP of 0.531. At last, the CARS-LS-SVM model was transferred to each pixel of hyperspectral images of samples and colour maps were generated for visualising the LAB spoilage process in salmon flesh. The overall results indicated that NIR hyperspectral imaging is very potential and could be used as a rapid, non-destructive and efficient technique for LAB evaluation in salmon flesh.     

Embedded bone fragment detection in chicken fillets using transmittance image enhancement and hyperspectral reflectance imaging

This paper is concerned with the detection of bone fragments embedded in compressed de-boned skinless chicken breast fillets by enhancing single-band transmittance images generated by back-lighting and exploiting spectral information from hyperspectral reflectance images. Optical imaging of chicken fillets is often dominated by multiple scattering properties of the fillets. Thus, resulting images from multiple scattering are diffused, scattered and low contrast. In this study, a fusion of hyperspectral transmittance and reflectance imaging, which is a non-ionized and non-destructive imaging modality, was investigated as an alternative method to the conventional transmittance X-ray imaging technique which is an ionizing imaging modality. An image formation model, called an illumination–transmittance model, was applied for correcting non-uniform illumination effects so that embedded bones are more easily detectable by a simple segmentation method using a single threshold value. Predicted bones from the segmentation were classified by the nearest neighbor classifier that was trained by the spectral library of mean reflectance of chicken tissues like fat, meat and embedded bones. Experimental results with chicken breast fillets and bone fragments are provided.