Stewart Postharvest Review

Stewart Postharvest Review An international journal for reviews in postharvest biology and technology

Spectral imaging techniques for food quality evaluation Ning Wang1*, Gamal ElMasry1, 2, S Sevakarampalayam1 and Jun Qiao1, 3 1

Department of Bioresource Engineering, McGill University, Ste-Anne-De-Bellevue, Quebec, Canada Faculty of Agriculture, Suez Canal University, Ismailia, Egypt 3 China Agricultural University, Beijing, PR China 2

Abstract Machine vision, spectroscopy and spectral imaging are potential fields of research for studying the external and internal properties of food products. This article reviews applications of these techniques in quality and safety evaluation, classification and sorting of food materials, and fruits and vegetables. Recent developments of imaging systems are discussed with emphasis on the use of multispectral and hyperspectral imaging for modern food inspection. Keywords: hyperspectral imaging; wavelength selection; multivariate statistical analysis; artificial neural network

Abbreviations ANN CCD CMOS FFN GLCM MIR MLR NIR PCA PFN PLS PSE r RFN RSE SEP VIS

Artificial Neural Network Charge-coupled Device Complementary Metal-oxide Semiconductor Feed-forward Neural Network Gray Level Co-occurrence Matrix Mid-infrared Multiple Linear Regression Near-infrared Principal Component Analysis Pale, Firm and Non-exudative Partial Least Square Pale Pinkish Gray, Very Soft and Exudative Correlation Coefficient Reddish Pink, Firm and Non-exudative Reddish, Soft and Exudative Standard Error of Prediction Visible

*Correspondence to: Ning Wang, Department of Bioresource Engineering, McGill University, Ste-Anne-DeBellevue, Quebec, Canada. Stewart Postharvest Review 2007, 1:1 Published online 01 February 2007 doi: 10.2212/spr.2007.1.1 © 2007 Stewart Postharvest Solutions (UK) Ltd. Online ISSN:1945-9656 www.stewartpostharvest.com

Introduction Technologies that can sort food products based on their appearance, texture, taste, flavour or nutritive value assure fruit quality and consistency, increase consumer confidence and satisfaction, and enhance the competitiveness and profitability of the fruit industry [1]. Currently, many food products are sorted manually or automatically depending on their external quality features. However, internal quality attributes such as dry matter content, total soluble solids content, sugar content and juice acidity are very important in modern quality evaluation industries. Most instrumental techniques for measuring these properties are destructive and involve a considerable amount of manual work. The development of nondestructive measurements for both external and internal quality attributes will be very useful for producers, processors and distributors to ascertain fast evaluation. Machine vision is an advanced technology used to “see” objects with the assistance of computers. It is capable of acquiring and quantifying an object’s external characteristics such as colour, size, shape and some textural attributes [2]. Hence, it has found many applications in pre- and postharvest product quality grading and sorting, eg, to detect the presence of diseases, defects, ripeness, maturity and other quality attributes in fruits and vegetables, grain and animal products. Visible (VIS)/near-infrared (NIR) spectroscopy is also recognised as an efficient and effective method for food quality assessment, with its capability of extracting the chemical information of food products. It relates the

Wang et al. / Stewart Postharvest Review 2007, 1:1

changes in VIS/NIR spectral responses of a food substance to the concentration changes of corresponding chemical components. Successful applications have be reported for predicting chemical compounds in various food products, eg, to predict the nutritive quality of corn silage [3], to predict moisture, protein and fat content of taramosalata, a famous Greek dish [4], to evaluate longissimus quality traits of beef carcasses [5], and to predict various quality parameters of wheat [6]. Standing on the shoulder of both traditional imaging and spectroscopic techniques, spectral imaging originated from the fields of chemistry and remote sensing has been widely used for assessing the quality aspects of agricultural produce [7]. With the development of optical sensors, spectral imaging integrates spectroscopy and imaging techniques to provide spectral information which has been implemented in applications such as inspection of poultry carcasses [8–10], defect detection or quality determination of fruits and vegetables [11–16], and estimation of physical, chemical and mechanical properties in various commodities [12, 17, 18]. Traditional machine vision, spectroscopic and hyperspectral imaging systems have many advantages over classical chemical and physical analytical methods; as they have a short measuring time with limited sample preparation, are chemical-free, and can be applied to estimate more than one attribute at the same time [19]. All these factors reduce energy requirements and process costs, and provide more consistent food to consumers. Traditional imaging systems acquire images from an object at three visible wavelengths of red (R), green (G) and blue (B). Little information on the internal property of an object can be found. The spectroscopic method has a great drawback because it acquires the spectral data from a single point or from a small portion of the tested fruit. Spectral imaging has advantages of receiving spatially distributed spectral responses at each pixel of a fruit image. A spectral image is composed of several sub-images, each representing the intensity of a certain spectral band. It integrates spectroscopy and traditional imaging techniques to provide spectral as well as spatial information for the surface of interest on a target object. Therefore, the spectral imaging system allows flexible selection of a region of interest, which improves the acquisition of spectral responses and then improves the accuracy of predicting the object’s attributes.

Spectral imaging systems To simultaneously acquire spatial and spectral information from a target object, each spectral imaging system consists of a subsystem to locate each pixel in an image (spatial) and a subsystem to provide spectral responses at predefined wavelengths at each pixel. Various configurations can be found for different spectral imaging systems. The two main categories are multispectral and hyperspectral imaging systems. Multispectral imaging systems A multispectral imaging system acquires spectral images by collecting a set of images at discrete wavebands using band-

pass filters with the predefined central wavelengths attached in front of the camera lens. Rotating filter wheels (manually or automatic controlled), with a fixed number of selected filters, are widely used due to their low cost and easy operation and maintenance. The liquid crystal and acousto-optic tunable filters can be electronically controlled to select a transmitted wavelength range while blocking all others, hence, providing rapid, vibrationless and smooth wavelength selection in the visible to infrared range. A more efficient approach is to use a common aperture multichannel imaging camera that simultaneously acquires multiple nonoverlapping images at selected wavebands. Lu [17] predicted the firmness and soluble solid content of red delicious apple fruit using multispectral imaging by adopting a spectral range of 680–1,060 nm for evaluation. Results showed that the spectral responses at 880, 905 and 940 nm could be used to predict the soluble solid content with a correlation coefficient (r) of 0.77 and a standard error of prediction (SEP) of 0.78%. NIR and mid-infrared (MIR) dual-camera imaging was used by Cheng et al. [16] for online detection of apple stem-end/calyx detection. This study was carried in apple out to distinguish the stem-end/ calyx from a true defect, which is useful in apple defect sorting systems. The algorithm developed to process the NIR and MIR images, when adopted for online tests, showed a 100% recognition rate for good apples and a 92% recognition rate for defective apples. Tran and Grishko [20] studied the water removal kinetics for olive leaves using an NIR multispectral imaging spectrometer. The imaging spectrometer used was an NIR camera with an acousto-optic tunable filter and an InGaAs focal plane array that has high sensitivity and is capable of rapidly recording NIR spectral images of leaves suitable for kinetic determination. The spectrum range adopted for their study was 1,000–1,700 nm. The results from the study for kinetics of water desorption showed that the rate of water desorption is strongly dependent on the environment in which the leaves are stored. Hyperspectral imaging systems Hyperspectral image acquisition is commonly based on a pushbroom method in which a line of spatial information with a full spectral range (visible to infrared) per spatial pixel is captured sequentially to complete a volume of spatialspectral data. Both complementary metal-oxide semiconductor (CMOS) and charge-coupled device (CCD) detectors with two-dimensional spatial data have been combined with a spectrograph which provides one-dimensional spectral data to create a three-dimensional spatial-spectral data cube. Due to the line-scan nature of the spectrograph, some hyperspectral imaging systems (eg, the AutoVision Hyperspectral Imaging System [AutoVision Inc., CA, USA]) need the target object to be moved by a conveyor in order to create the second spatial dimension, while other systems (eg, the Optikon VNIR 100E hyperspectral imaging system [Optikon Corp., ON, Canada]) move the camera lens with a special optical mechanism that is more flexible and portable. 2

Wang et al. / Stewart Postharvest Review 2007, 1:1 Figure 1. The hyperspectral imaging system for pork meat grading [37].

Abbreviation: CMOS, complementary metal-oxide semiconductor

Tatzer et al. [21] developed an industrial online material sorting system with the spectral imaging technique. Functional components and classification methods were studied for cellulose-based materials such as pulp, paper and cardboard. Patrick et al. [22] developed a hyperspectral imaging technique with a high spatial resolution (0.5–1.0 mm). The system was able to detect defective and contaminated foods and agricultural products. A hyperspectral imaging system was also successfully developed for inspecting the contamination of chicken carcasses [10, 23]. Nagata et al. [18] adopted the hyperspectral imaging technique predict the maturity of strawberries. The authors developed prediction models to estimate firmness and soluble solids content in strawberry in the visible range of 450–650 nm. The results indicated that individual maturity level analysis was necessary for adequate prediction of soluble solids content. Lu [24] investigated the potential of NIR hyperspectral imaging for detecting bruises in apple in the spectral range 900–1,700 nm. The study involved detection of bruises on apples at different days after bruising. The author concluded that the spectral region between 1,000 and 1,340 nm was most appropriate for detecting bruises in apple. Kim et al. [25] developed a laboratory-based hyperspectral imaging system with a spectral range of 430–930 nm at a spectral resolution of approximately 10 nm to be used for research in food quality and safety. The system was tested and was shown to be useful for sampling the fluorescence and reflectance images of normal, contaminated and bruised/ defective apples. Cheng et al. [26] used a hyperspectral imaging system to extract feature bands for detecting chilling injury in cucumber. Guyer and Yang [27] collected spectral images for cherries over the 680–1,280 nm range at increments of 40 nm. The study resulted in an average of 73% classification accuracy for correct identification as well as quantification of all types of cherry defects.

Spectral image processing and data analysis Data volume reduction and optimal wavelength selection The major disadvantage of hyperspectral imaging is that handling the large amount of data extracted from hyperspectral images requires extra time and resources. It is imperative that efficient manipulation of procedures be used to reduce data dimensionality to its lowest level without losing functionality. Instead of the entire image volume (x × y × n, x-spatial dimension, y-spatial dimension, n-spectral dimension) being used, a reduced data cube with dimensions of x×y × ns, where ns is the number of selected optimal wavelengths, was formed. The success of a classification algorithm based on the reduced data cube depends on the quality of the selection of the optimal wavelengths at which the spectral signatures can best describe each class. Several wavelength selection techniques have been derived and include general visual inspection of the spectral curves and correlation coefficients [28], analysis of spectral differences from the average spectrum [29], stepwise regression [30], principal component analysis (PCA) [13, 31], and other techniques [32]. Averaging the spectrum with certain bandwidth (eg, every 5–10 nm) can be applied for this task but some important information that might be carried by certain bands could be lost. Xing and Baerdemaeker [33] constructed a hyperspectral imaging system with a wavelength range of 400–1,000 nm for detecting bruises on ‘Jonagold’ apples. PCA was used to visualise the hyperspectral data and develop algorithms for the selection of efficient wavebands. Based on the algorithms for bruise detection and the stem-end/calyx identification the correct classification rate for sound apples was 84.6 and 77.5% for 1day-old bruises. The stem-end/calyx presented in the images was recognised to the accuracy of 98.3%, with only 2.5% of bruises misclassified as stem-end/calyx. 3


Figure 2. Masking operations: (a) band image at 685 nm; (b) image after binarisation using histogram valley method; (c) image after multiple erosions; (d) image after filling holes and erosion again; (e) image mask obtained after particle removal [37].

(a)

(b)

Artificial neural network (ANN) models have been developed to simulate some of the organisational principles from observations of the human brain and nervous system. ANNs have the advantage of solving problems in which some inputs and corresponding output values are known, but the relationship between the inputs and outputs is not well understood or is difficult to translate into a mathematical function. Despite mathematical differences, ANNs have shown strength over statistical classifiers in identifying and classifying agricultural produce [34], where non-coherence or non-linearity often exists. ANNs have been used for data volume reduction and wavelength selection on the basis of the fact that the network can change and adjust its knowledge by adjusting its weights according to the presented samples of data.

Parameter prediction and classification models Machine vision, VIRS/NIRS and hyperspectral imaging techniques require mathematical procedures to establish models for evaluating the internal properties of the object. Samples with known composition, called training sets, are used to establish the relationship between the spectral information and the chemical constituents. The typical algorithms used by various researchers are PCA, partial least square (PLS), ANN, multiple linear regression (MLR), and other multivariate statistical analyses for classification and prediction. VIS and NIR spectroscopy in the spectral range of 400–2,500 nm was used by Cozzolinoa and Murray [35] to identify and authenticate different meat muscle species. PCA models and dummy PLS models were developed for identifying different meat species and for developing predictive equations. The feasibility of an NIR hyperspectral imaging spectrometer application for the quality analysis of single maize kernel was tested by Cogdill et al. [36]. Qiao et al. [37] established a hyperspectral imaging system to grade pork meat. Two sets of feed-forward neural network (FFN) prediction models were developed to predict drip loss, pH and colour of pork meat with r values of 0.77, 0.55 and 0.86, respectively.

(c)

(d)

(e)

Case study Prediction of pork quality attributes using hyperspectral imaging technique The water holding capacity of pork meat indicates its aptitude for further processing. Statistical data show that exudative pork meats can induce an economic loss of US $5 per carcass [38]. Meats are often graded into four classes: (1) reddish pink, firm and non-exudative (RFN), with a desirable colour, firmness, normal water-holding capacity, minimal drip loss and moderate decline rate of pH; (2) pale pinkish gray, very soft and exudative (PSE), with undesirable appearance and excessive shrinkage, very poor water-holding capacity, excessive drip loss and a rapid decline rate of pH (5.6–5.5); (3) reddish, soft and exudative (RSE), with normal colour, but a soft texture and an exudative character similar to PSE [39]; (4) and pale, firm and non-exudative (PFN) [40]. Rapid and objective techniques are urgently needed for characterising exudative properties of pork meat. A hyperspectral imaging system consisting of a line-scan spectrograph (ImSpector, V10E, Spectra Imaging Ltd, Finland), a CMOS camera (BCi4-USB-M40LP, Vector International, Belgium), a DC illuminator (Fiber-Lite PL900-A, Dolan-Jenner Industries Inc, USA), a conveyer (Dorner 2200 series, Donner Mfg. Corp., USA), an enclosure, data acquisition and preprocessing software (SpectraCube, Auto Vision Inc., USA), and a computer was established as shown in Figure 1. Simple correlation analysis was conducted between the spectral response at each wavelength from 430–980 nm and corresponding drip loss, pH and colour. The wavelengths at which the highest correlation coefficient was found were selected as feature wavebands. The feature band images were then extracted from original spectral images at the feature wavebands. The feature band image at 685 nm was used to make a mask due to its significant contrast between the loineye area and the background. With multiple processes of erosion and filling-holes, the loin-eye area and the other small 4


Figure 3. Average spectra and its first derivative of four meat quality classes (RFN, PFN, PSE and RSE) [37].

Abbreviations: PFN, pale, firm and non-exudative; PSE, pale pinkish gray, very soft and exudative; RFN, reddish pink, firm and non-exudative; RSE, reddish, soft and exudative

lean area around the loin were extracted (Figure 2). The average spectra and its first derivative of four meat quality classes extracted from the marked area are shown in Figure 3. Six feature band images were selected for predicting the drip loss (655, 685, 755, 618, 459 and 953 nm), pH (669, 978, 571, 637, 703 and 494 nm) and colour (434,494,561, 637,669 and 703 nm). The average intensity of the band images was used to predict the quality attributes. FFN models were developed. Results showed that the FFN model could predict the drip loss, pH and colour of the meat sample with r values of 0.77, 0.55 and 0.86, respectively. Subgroups of exudative and non-exudative, reddish and pale meats were successfully classified using predicted drip loss and pH.

consists of two 50 W halogen lamps adjusted at angle of 45º to illuminate the camera’s field of view; (2) a fruit holder surrounded by a cube tent made from white nylon to diffuse the light and provide an optimum lighting condition; (3) a spectrograph (ImSpector V10E, Optikon Co., Canada) coupled with a standard C-mount zoom lens; and (4) a CCD camera (PCO-1600, PCO. Imaging, Germany). The assembly disperses the incoming line of light into the spectral and spatial matrices and then projects them onto the CCD. The optics, spectrograph and the camera have high sensitivity from 400 to 1,000 nm and the exposure time was adjusted to 200 ms throughout the entire test. The distance between the lens and the surface of the strawberry being imaged was fixed at 40 cm. After finishing the scans on a fruit, a threedimensional (x, y, λ) spatial and spectral data space was constructed. Images were binned during acquisition in spatial direction to provide images with a spatial dimension (x, y) of 400×400 pixels, with 826 spectral bands (λ) of a spectral resolution of 0.73 nm. The hyperspectral imaging system was controlled by a computer with Hypervisual Imaging Analyzer software (ProVision Technologies, Stennis Space Center, USA) for image acquisition. PLS analysis was applied to build the prediction model of quality attributes. PLS analysis between one attribute (moisture content, soluble solids or pH) and the spectral data (average spectra with 826 wavelengths in the range from 400 to 1,000 nm) was conducted. To avoid possible loss of important information during wavelength selection, spectral responses on all 826 wavebands were used without variable reduction process. β-coefficients resulting from the best PLS calibration model were used for identifying the optimal wavelengths. The wavelengths that corresponded to the high-

Figure 4. Hyperspectral imaging system: (a) a CCD camera; (b) a spectrograph with a standard C-mount zoom lens; (c) a halogen lighting unit; (d) a white nylon tent; and (e) a computer supported with an image-acquisition software [41].

Nondestructive determination of quality attributes for strawberry Strawberry (Fragaria sp.) is one of the economically important fruits that are more popularly eaten fresh, used for garnishing cakes and pastries, used as flavouring for juices and milk products, and processed into jams and others products. Together with the recent concern for food quality and safety, automatic technologies for judging the quality of fresh strawberry are being sought. A laboratory hyperspectral imaging system was constructed by Qiao et al [37] as shown in Figure 4. It is composed of the following four components: (1) an illumination unit which

Abbreviation: CCD, charge-coupled device 5


Table 1. Performance of MLR models for predicting moisture content, total soluble solids and pH using only the optimal wavelengths extracted from β-coefficients of PLS analysis.

Attribute

MC

Optimal wavelengths

480, 528, 608, 685, 753, 817, 939, 977

Calibration

Validation

SEC

r

SEP

r

6.72

0.87

5.786

0.91

0.220

0.80

0.211

0.80

0.084

0.92

0.091

0.94

421, 520, 581, TSS

pH

683,847, 950

421, 521, 585, 646, 681, 840, 950, 990

Abbreviations: MC, moisture content; MLR, multiple linear regression; PLS, partial least square; r, correlation coefficient; SEC, standard error of calibration; SEP, standard error of prediction; TSS, total soluble solids

classified into three ripeness categories: unripe, ripe and overripe. In this study, fruits containing a green area equal or more than one third of its surface were classified as unripe fruits. The classification of the fruits in these categories was conducted depending on the visual appearance of the fruit. A RGB image was constructed for each fruit by picking up the red (650 nm), green (500 nm) and blue (450 nm) bands from the corrected hyperspectral image to form a colour image. This colour image was the basis for extracting the texture parameters. Texture measures were derived from the gray level co-occurrence matrix (GLCM), which is a square matrix with elements corresponding to the relative frequency (Pi,j) of occurrence of pairs of gray levels of pixels separated by a certain distance (D) in a given direction (0, 45, 90 or 135°) as shown in Figure 5d. The discrimination efficiency at different directions of 0, 45, 90 and 135° were found to be 89.61, 77.62, 81.82 and 79.22 %, respectively. It is clear that the horizontal direction at angle 0° had the highest discrimination efficiency (89.61%) compared with other directions. Only the confusion matrix at this direction for fruit classification is shown in Table 2.

Summary est absolute values of β-coefficients were considered optimal wavelengths. Wavelengths which corresponded to the lowest absolute values of β-coefficients were completely neglected because they have no or very little contribution in prediction. Only the selected optimal wavelengths were used to establish multiple linear regression prediction models. Table 1 shows the selected wavelengths and results from MLR analysis. Texture analysis was conducted to identify the ripeness stage of strawberry [41]. As shown in Figure 5, tested fruits were

Machine vision technology has shown great potential in many agri-food applications. Advantages of using machine vision systems as sensors include their high resolution and non-destructive nature. Specialised machine vision systems can see objects beyond the visible colour region, reaching the invisible ultraviolet, near-infrared or infrared wave bands. Information received from the expanded wave bands can be very useful in determining many properties, including variety, defects, maturity, ripeness, diseases, contaminations and other important quality/safety indicators. Hyperspectral imaging techniques combine conventional, two-dimensional

Figure 5. RGB image for (a) unripe, (b) ripe and (c) overripe strawberries. The cropped small square for texture analysis is depicted at the bottom left corner of each image. (d) Extraction of GLCM at different directions (0, 45, 90, and 135°) and distance (D) for each pixel in the cropped square image [41].

135 ° [-D -

90 ° [-D 0]

Pixel of interest

50

45 ° [-D D]

0 ° [0 D]

50 (a)

(b)

(c)

(d)

Abbreviation: GLCM, gray level co-occurrence matrix 6


Table 2. Confusion matrix for ripeness classification using discriminant analysis with cross-validation at 0° direction. From/to

Unripe

Unripe

7

Ripe

Ripe

Overripe

Total

% Correct

2

0

9

77.78

0

14

5

19

73.68

Overripe

0

1

48

49

97.96

Total

7

17

53

77

89.61

7

8

9

10

11

digital imagery with spectroscopy to detect even minor and subtle spectral features at an extremely high spatial resolution. Thus, hyperspectral imaging can be directly used as a sensor and as a fundamental tool to study the properties of agri-food products, as well as to extract feature signatures for properties of interest. These feature signatures can then be used as a base to develop cost-effective multispectral sensors that use combined spectroscopic and image-processing algorithms for detection and inspection. As mentioned before, the major disadvantage and barrier for multi- and hyperspectral imaging to be used for fast, realtime food quality assessment is their requirement of time and resource to process the huge three-dimensional data cube. With the development of ultra-high-speed microprocessors, low-cost, high-performance memory and new data processing algorithms, spectral imaging will soon find its applications in online food quality grading.

References Papers of interest have been highlighted as: *Marginal importance **Essential reading

12

13

Kim MS, Lefcourt AM, Chao K, Chen YR, Kim I and Chan DE. Multispectral detection of fecal contamination on apples based on hyperspectral imagery: part I. Application of visible and near–infrared reflectance imaging. Transactions of the ASAE 2002. 45(6): 2027–2037.

14

Polder G, Van der Heijden GWAM and Young IT. Spectral image analysis for measuring ripeness of tomatoes. Transactions of the ASAE 2002: 45: 1155–1161.

15

Liu Y, Chen YR, Wang CY, Chan DE and Kim MS. Development of hyperspectral imaging technique for the detection of chilling injury in cucumbers; spectral and image analysis. Applied Engineering in Agriculture 2006: 22(1): 101–111. Cheng X, Tao Y, Chen YR and Luo Y. NIR/MIR dual–sensor machine vision system for online apple stem–end/calyx recognition. Transactions of the ASAE 2003 46(2): 551–558. Lu R. Multispectral imaging for predicting firmness and soluble solids content of apple fruit. Postharvest Biology and Technology 2004: 31(1): 147–157.

16

17

18

Nagata M, Tallada JG, Kobayashi T and Toyoda H. NIR hyperspectral imaging for measurement of internal quality in strawberries. ASAE Paper No. 053131. ASAE, St . Joseph, MI, USA; 2005.

19

Lammertyn J, Nicolaï B, Ooms K, De Smedt V and De Baerdemaeker J. Non-destructive measurement of acidity, soluble solids, and Firmness of Jonagold apples using NIR-spectroscopy. Transactions of the ASAE 1998: 41(4): 1089–1094. Tran CD and Grishko V. Determination of water contents in leaves by a near infrared multispectral imaging technique. Microchemical Journal 2004: 76: 91–94. Tatzer P, Wolf M and Panner T. Industrial application for inline material sorting using hyperspectral imaging in the NIR range. Real-Time Imaging 2005: 11: 99–107. Mehl Patrick M, Chen Yud-Ren, Kim Moon S and Chan Diane E. Development of hyperspectral imaging technique for the detection of apple surface defects and contaminations. Journal of Food Engineering 2004: 61: 67–81.

1

Lu R and Ariana D. A near-infrared sensing technique for measuring internal quality of apple fruit. Applied Engineering in Agriculture 2002: 18(5): 585–590.

20

2

Chen YR, Kuanglin Chao and Kim Moon S. Machine vision technology for agricultural applications. Computers and Electronics in Agriculture 2002: 36: 173–191.

21

3

Lovett DK, Deaville ER, Givens DI, Finlay M and Owen E. Near infrared reflectance spectroscopy (NIRS) to predict biological parameters of maize silage: effects of particle comminution, oven drying temperature and the presence of residual moisture. Animal Feed Science and Technology 2005: 120: 323–332. Adamopoulos KG and Goula AM. Application of near-infrared reflectance spectroscopy in the determination of major components in taramosalata. Journal of Food Engineering 2004: 63: 199–207. Shackelford SD, Wheeler TL and Koohmaraie M. Development of optimal protocol for visible and near-infrared reflectance spectroscopic evaluation of meat quality. Meat Science 2004: 68: 371–381.

4

5

6

Dowell FE, Maghirang EB, Xie F, Lookhart GL, Pierce RO, Seabourn

BW, Bean SR, Wilson JD and Chung OK. Predicting wheat quality characteristics and functionality using near-infrared spectroscopy. Cereal Chemistry 2006: 83(5):529–536. Kavdir I and DE Guyer. Apple sorting using artificial neural networks and spectral imaging. Transactions of the ASAE 2002: 45(6): 1995– 2005. Chao K, Chen YR, Hruschka WR and Park B. Chicken heart disease characterization by multi-spectral imaging. Transactions of the ASAE 2001: 17: 99–106. Park B, Windham WR, Lawrence KC and Smith DP. Hyperspectral image classification for fecal and ingesta identification by spectral angle mapper. ASAE Paper No. 043032. ASAE/CSAE Meeting, Ottawa, Ontario, Canada; 2004. Kim I, Kim MS, Chen YR and Kong SG. Detection of skin tumors on chicken carcasses using hyperspectral fluorescence imaging. Transactions of the ASAE 2004: 47(5):1785–1792. Li Q, Wang M and Gu W. Computer vision based system for apple surface defect detection. Computers and Electronics in Agriculture 2002: 36 (2): 215–223. Park B, Abbott JA, Lee KJ, Choi CH and Choi KH. Near-infrared diffuse reflectance for quantitative and qualitative measurement of soluble solids and firmness of Delicious and Gala apples. Transactions of the ASAE 2003: 46(6): 1721–1731.

22

23

Yang CC, Chao K and Chen YR. Development of multispectral image processing algorithms for identification of wholesome, septicemic, and inflammatory process chickens. Journal of Food Engineering 2005: 69: 225–234.

24

Lu R. Detection of Bruises on Apples using Near-Infrared Hyperspectral Imaging. Transactions of the ASAE 2003: 46(2): 523–530. Kim MS, Chen YR and Mehl PM.. Hyperspectral reflectance and fluorescence imaging system for food quality and safety. Transactions of the

25

7

Wang et al. / Stewart Postharvest Review 2007, 1:1 ASAE 2001: 44(3): 721–729. 26

Cheng X, Chen YR, Tao Y, Wang CY, Kim MS and Lefcourt AM. A novel integrated PCA and FLD method on hyperspectral image feature extraction for cucumber chilling damage inspection. Transactions of the ASAE 2004: 47(4), 1313−1320.

27

Guyer D and Yang X. Use of genetic artificial neural networks and spectral imaging for defect detection on cherries. Computers and Electronics in Agriculture 2000: 29: 179–194.

28

Keskin M, Dodd RB, Han YJ and Khalilian A. Assessing nitrogen content of golf course turfgrass clippings using spectral reflectance. Applied Engineering in Agriculture 2004: 20(6): 851−860.

29

Liu Y, Windham WR, Lawrence KC and Park B. Simple algorithms for the classification of visible/near-infrared and hyperspectral imaging spectra of chicken skins, feces, and fecal contaminated skins. Applied Spectroscopy 2003: 57(12): 1609–1612. Chong L-G and Jun C-H. Performance of some variable selection methods when multicollinearity is present. Chemometrics and Intelligent Laboratory Systems 2005: 78(1): 103–112. Mehl PM, Chen Y-R, Kim MS and Chan DE. Development of hyperspectral imaging technique for the detection of apple surface defects and contaminations. Journal of Food Engineering 2004: 61(1): 67–81. Hruschka WR. Near infrared technology in the agricultural and food industries. Data analysis: wavelength selection methods, Edited by Phil Williams and Karl Norris. American Association of Cereal Chemists, St Paul, MN, USA; 2001. Xing J and De Baerdemaeker. J Bruise detection on ‘Jonagold’ apples using hyperspectral imaging. Postharvest Biology and Technology 2005:

30

31

32

33

37(1): 152–162. 34

Jayas DS, Paliwal J and Visen NS. Multi-layer neural networks for image analysis of agricultural products. Journal of Agricultural Engineering Research 2000:77(2): 119–128.

35

Cozzolino D and Murray I. Identification of animal meat muscles by visible and near infrared reflectance spectroscopy. LebensmittelWissenschaft und Technologie 2004: 37: 447–452.

36

Cogdill RP, Hurburgh Jr CR and Rippke GR. Single-kernel maize analysis by Near-infrared hyperspectral imaging. Transactions of the ASAE 2004: 47(1): 311–320.

37

Qiao J, Wang N, Ngadi MO, Gunenc A, Monroy M, Gariépy C and Prasher SO. Prediction of drip-loss, pH, and color for pork using hyperspectral imaging technique. Meat Sciences (In Press).

38

Murray AC. Reducing losses from farm gate to packer: A Canadian’s perspective. Proceedings of the 1st International Virtual Conference on Pork Quality. Concordia, Brazil; 2000: pp. 72–84.

39

Kauffman RG, Cassens RG, Scherer A and Meeker DL. Variations in pork quality. National Pork Producer Council, Des Moines, IA, USA; 1992.

40

Nam KC, Du M, Jo D and Ahn U. Effect of ionizing radiation on quality characteristics of vacuum-packaged normal, pale-soft-exudative, and dark-firm-dry pork. Journal of Innovative Food Science and Emerging Technologies 2002: 3:73–79.

41

ElMasry G, Wang N and Ngadi MO. Hyperspectral imaging for nondestructive determination of quality attributes for strawberry. Journal of Food Engineering (In press).

© 2007 Published by Stewart Postharvest Solutions (UK) Ltd. All rights reserved. www.stewartpostharvest.com 8