Author’s Accepted Manuscript Rapid detection of frozen pork quality without thawing by Vis-NIR hyperspectral imaging technique Anguo Xie, Da-Wen Sun, Zhongyue Xu, Zhiwei Zhu www.elsevier.com
PII: DOI: Reference:
S0039-9140(15)00109-5 http://dx.doi.org/10.1016/j.talanta.2015.02.027 TAL15400
To appear in: Talanta Received date: 29 October 2014 Revised date: 6 February 2015 Accepted date: 17 February 2015 Cite this article as: Anguo Xie, Da-Wen Sun, Zhongyue Xu and Zhiwei Zhu, Rapid detection of frozen pork quality without thawing by Vis-NIR hyperspectral imaging technique, Talanta, http://dx.doi.org/10.1016/j.talanta.2015.02.027 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting galley proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
Rapid detection of frozen pork quality without thawing by Vis-NIR hyperspectral imaging technique Anguo Xiea, Da-Wen Suna,b, Zhongyue Xua, Zhiwei Zhua
a
College of Light Industry and Food Sciences, South China University of Technology, Guangzhou, 510641, China
b
Food Refrigeration and Computerised Food Technology, University College Dublin, National University of Ireland, Agriculture and Food Science Centre, Belfield, Dublin 4, Ireland
Abstract Quality determination of frozen food is a time-consuming and laborious work as it normally takes a long time to thaw the frozen samples before measurements can be carried out. In this research, a rapid and non-destructive determination technique for frozen pork quality was tested with a hyperspectral imaging (HSI) system. In this study, 120 pieces of pork meat were frozen by four kinds of methods with various freezing temperatures from -20 to -120 °C. The hyperspectral images of the samples were acquired at the frozen state. Quality indicators including drip loss, pH value, color, cooking loss and Warner-Bratzler shear force (WBSF) of the samples were measured after thawing. The spectral characteristics of the frozen meat samples were studied and it was revealed that the reflectance at 1100 nm had a close relationship with the freezing temperature (R = -0.832, p < 0.01). Partial least squares regression (PLSR) was applied to establish the spectral models, and the models were then optimized. Results showed that the improved region of interest (ROI) method could be used to extract effective spectral information to withstand the interference of freezing, and choosing appropriate spectral bands and spectral pretreatment techniques were crucial to develop robust mathematical model. The performances of the models established were diverse based on different quality indicators. The coefficients of determination for prediction 2
( R p ) for L*, cooking loss, b*, drip loss and a* were 0.907, 0.845, 0.814, 0.762, and 0.716, respectively. However 2
there were low correlations ( R p ) for pH and WBSF measurements. The current study indicated that HSI had the potential for non-destructive determination of frozen meat quality without thawing.
-1-7167342, Fax: +353-1-7167493, E-mail: [email protected],
Website: www.ucd.ie/refrig; www.ucd.ie/sun.
Keywords: Hyperspectral imaging; frozen food; freezing; NIR spectroscopy; drip loss; meat color
1. Introduction Pork is one of the most valuable meat products and is the primary choice for people to obtain protein. However pork meat is perishable [1], therefore freezing is a common technique to maintain its quality. When meats are in frozen state, it is not easy to determine their quality. Therefore the frozen meat market is mixed with superior and inferior products. It was reported that some food companies froze expired or inferior meats and sold them [2]. In addition, different freezing techniques can also have influences on the quality of frozen meats [3-4]. For example, slow freezing can promote the growth of large ice crystals, which could damage cell structure, upon thawing, it can cause more loss of meat juices
[5-6]
and color fading or darkening. Furthermore, temperature
fluctuations during refrigeration storage can also deteriorate the quality of the frozen foods, causing protein denaturation and fat oxidation [7], and spoilage [8-9]. Therefore, it is essential to ascertain the quality of frozen food products. Many indicators could be used to characterize the quality of frozen pork meats including drip loss after thawing, cooking loss, pH, color, and tenderness (Warner-Bratzler shear force). Among them, drip loss and cooking loss are two important indexes that have direct impact on the economic benefits of a company. Traditional methods are used to measure these quality indicators. Drip loss and cooking loss are measured by weighing the thawed or cooked sample and then comparing the weight with the initial weight of the frozen samples. In measuring drip loss, samples are usually thawed for 12 or 24 hours
[10]
. The pH value can be
determined by a pH meter, and its changes with storage time could indicate the freshness of the samples
[11-12]
.
Pork color can be measured by a colorimeter, which impacts on sensory evaluation and consumer desires. Tenderness reflects the palatability of the products, and can be determined by Warner-Bratzler shear force. All these measuring techniques are time-consuming and in particular, cannot meet the requirement of on-line measurements. Although some rapid detection studies have been carried out, for example Hardy et al. [13] detected the microbial contamination in frozen vegetables by automated impedance measurements and Koch et al. [14] used rapid polymerase chain reaction method to detect vibrio cholerae in foods, these studies focused on only microbiological detection, especially, it still needed a long time (more than 5 h) for testing. Therefore rapid and
2
non-destructive detection technique for more quality indicators is needed. Hyperspectral imaging (HSI) is a novel non-destructive evaluation method, which has been widely investigated for applications in the agricultural and food industry
[15]
. HSI provides both spatial and spectral
information for each pixel in the image. Therefore HSI can not only capture the external attributes (size, color, shape, surface texture, etc.) as traditional imaging technology, but can also be able to identify the chemical composition
[16]
in food as spectroscopic technique. Many studies have been conducted on using HSI for
determining meat quality. Elmasry et al.
[17]
developed a near infrared (NIR) HSI (900–1700 nm) system for the
measurement of surface color, pH and tenderness of fresh beef, Li et al.
[18]
used 400-1100nm HSI to assess
beef-marbling grade and obtained a good result with R2 = 0.92. Similar techniques were also employed to detect microbial contamination of porcine meat
[19]
, chicken contamination
[20]
, lamb meat quality
[21]
and so on. In
particular, Barbin et al. [22] evaluated the fresh and frozen–thawed porcine longissimus dorsi muscles by NIR HSI. Similar study was also reported by Cheng et al.
[23]
for fish fillets. However in these studies, the frozen samples
were thawed and then their images were acquired at room temperature by HSI. To the best of our knowledge, no research is available on using HSI technique to acquire food images at frozen state without thawing, and directly evaluate the frozen food quality. Therefore, the current study paper aimed to study a rapid and non-destructive technique based on hyperspectral imaging for direct measurement of frozen pork meat quality. The study also aimed to extract the spectral features of the frozen samples and to evaluate the impact of ice and frost on spectral detection.
2. Materials and methods The main steps of this experiment are shown in Fig. 1, which included freezing pork meat samples, image acquisition, measurement of quality indicators, spectral analysis, modeling and optimizing, and visualization. The experimental procedures are detailed below. 2.1. Freezing pork samples Longissimus dorsi meat samples from ten pigs were obtained from a local market at 0.5 day post-mortem. Then the pork samples were cut into 120 pieces (4 cm x 5 cm x 10 cm) with mass of 200 ± 4 g. These pork samples were frozen by four kinds of techniques including cryogenic freezing using liquid nitrogen, immersion freezing, air-blast freezing and domestic freezing in a conventional refrigerator. The main parameters of the
3
freezing processes are presented in Table 1. T-type thermocouple was inserted into the center of the sample to record the temperature variation with a data logger (TC-08, Pico Technology, Cambridgeshire, UK). There are a total of 120 samples, and each group of experiments contained 5 samples. When the core temperature of all 5 samples of one group reached -20 °C, the samples were moved into a refrigerator (BL/BD-719H, Haier Ltd., Qingdao, China) to store for 3 days at -20 °C. 2.2. Hyperspectral imaging system Spectral images were acquired by a hyperspectral imaging system shown in Fig. 2, which included a visible (Vis) unit (400 - 1000 nm) and a NIR unit (900 - 2500 nm). The Vis hyperspectral imaging system consisted of a camera lens (OLE23, Schneider, Rueil, German), a spectrograph (Imspector V10E, Spectral Imaging Ltd., Oulu, Finland) for the spectral range of 400–1000 nm, a high performance 1004 × 1002 charge coupled device (CCD) camera (DL-604 M, Andor, Belfast, Ireland), two 150 W halogen lamps (2900-ER, Illumination Technologies Inc., New York, USA) forming the illumination unit. The NIR unit consisted of a C-mount lens, a 12-bit CCD camera (XEVA 992, XC 130 XenICs, Leuven, Belgium) with 320 × 300 pixels, a spectrograph (ImSpector N17E, Specim, Oulu, Finland) covering the spectral range of 900–2500 nm, and an illumination unit of two 500 W tungsten halogen lamps (Lowel V-light™, NY, USA). The common parts of this Vis-NIR hyperspectral imaging system were a translation stage operated by a steppermotor (IRCP0076-1COMB, Isuzu Optics Co., Taiwan, China), and a computer supported with a data acquisition software (Spectral Image software, Isuzu Optics Co., Taiwan, China). 2.3. Image acquisition and calibration The difference of the Vis-NIR HSI system used in the current study from common hyperspectral imaging systems used in most studies was that a mini refrigerator (FYL-YS-30L, Fuyilian Co., Beijing, China) was installed on the translation stage. The temperature of the refrigerator was set at -20 °C. In the process of acquiring image, frozen pork sample was put in the refrigerator to avoid thawing. Every sample was scanned twice, i.e., the Vis image was acquired at the first, and then the NIR image was obtained by switching from the Vis system to the NIR system. In order to reduce the noise of the images, the acquired images were calibrated to obtain the calibrated image (R) using the following equation:
R
R 0 Rd 100% Rw Rd
(1)
4
where R0 is the raw hyperspectral image; Rd is the dark reference image of 0% reflectance, which was acquired by completely closing the lens of the camera with the opaque cap; and Rw is the white reference image (99%) obtained by using a white Teflon calibration tile [24]. 2.4. Measurement of quality parameters In this experiment, five quality indicators including drip loss, pH value, color value, cooking loss, and tenderness (Warner-Bratzler shear force) were measured. After image acquisition, the mass of the frozen pork samples (m1) was immediately weighed by an electronic balance (AL204, Mettler Toledo, Shanghai, China). Then the samples were thawed at the temperature of 4 °C for 24 hours
[5]
, the mass of the thawed samples (m2) was
measured again and the drip loss was calculated according to the equation below: Drip loss (%) =
m2 m1 100% m1
(2)
The pH value of the thawed sample was measured using a hand-held pH meter (Model 205, Testo AG, Lenzkirch, Germany), and the meat color was measured by a Chroma meter (CR-400, Konica Minolta Optics, Tokyo, Japan) in L*, a*, and b*. The measurement of the cooking loss was based on the procedure described by Honikel et al.
[10]
, i.e., the
frozen-thawed sample was cooked in a water bath at 80 °C until the core temperature of the sample reached 75 °C. The cooked sample was chilled at 4 °C for 4 h and then mass (m3) was taken by the balance. The cooking loss was calculated according to the following equation: Cooking loss (%) =
m3 m1 100% m1
(3)
where m1 is the mass of the frozen sample. Warner–Bratzler shear force (WBSF) is a recognized measurement for the tenderness of meats. In this experiment, every cooked sample was cut into three meat cuboids with a l00 mm2 (10 mm x 10 mm) cross-section area with the muscle fibre direction parallel to the long dimension of at least 30 mm
[10]
. Then the meat cuboids
were sheared with an Instron universal testing machine (5944S3013, Instron, Boston, USA) at the right angle to the fibre direction, and the shear force (N) was obtained to represent the tenderness. 2.5. Image segmentation and spectra extraction
5
In order to obtain useful spectral information, two kinds of methods were used to extract the spectra: (1) the background was removed from the hyperspectral image by subtracting a low-reflectance band from a high-reflectance band and the image was segmented by thresholding. All pixels of the frozen meat sample were selected as the region of interest (ROI), and then the spectra in ROI were averaged as the spectra of the whole sample
[25-26]
; (2) it was an improved ROI selecting method, in which not only the background but also the frost
region and shadow region were removed by using ENVI 4.8 software (Exelis Visual Information Solutions, Boulder, USA) as shown in Figure 3. The operation procedure in the 2nd method was to first open the ROI Tool in the ENVI software, then select preliminary ROI and finally press down "Grow" function key. The pixels in pre-existing ROI would be divided according to the deviation of spectra. 2.6. Establishment of spectral model and image visualization As one of the most common regression methods in chemometrics, partial least squares regression (PLSR) was used in the current study for predicting the quality indicators of the frozen pork meats. Among the 120 samples, they were randomly sorted, and then 90 samples were used as the calibration set and the remaining 30 samples were used for validation. The PLSR models were verified by cross-validation to confirm latent variables (LVs). The coefficients of determination for prediction ( R p2 ) and the root mean square error of prediction (RMSEP) were used as the criteria to assess the performance of the models. These algorithms were written and implemented in Matlab 2010b (MathWorks, Natick, USA). As a hyperspectral image is a three-dimensional matrix, which includes two-dimensional pixel matrix and one-dimensional spectral data, the spectra of each pixel were used as the input to the prediction models, and the result of each pixel was quantified. Then image visualization and prediction map were developed [13]. 3. Results and discussion 3.1. Quality indicators The quality indicators of the frozen pork samples are shown in Table 2 and Table 3. As indicated in Table 2, the mean value of pH for 120 samples was 5.61 ± 0.07. Many studies [11-12] have confirmed that pH of pork meat is greatly affected by storage time. For living pig, the pH of muscle is about 7.0. After slaughter, rigor mortis occurs and pH value decreases rapidly, and then pH increases slowly due to the ageing and spoilage of meat. In the current study, as all pork samples were stored for the same period after slaughtering, the pH values of the samples were similar. In addition, the data in Table 2 show that freezing and thawing did not affect the pH of the meat. 6
Meat color is an important indicator that can impact on consumer desires, and L*, a* and b* are used to indicate the brightness, red degree, and yellowness of the meat samples, respectively. Generally speaking, meat color is determined by the age of an animal, the amount of exercise it gets, the protein and fat content, and the storage period. Especially oxidation of myoglobin is a major factor for red color. As shown in Table 2, the average value of L*, a* and b* was 51.36, 3.99 and 5.62, respectively, and there existed strong correlations among the three color indicators of the meat samples, i.e., L* was significantly correlated with a* (R = -0.4, p < 0.01), and a* was highly correlated with b* (R = 0.571, p < 0.01). Table 2 also shows that the drip loss was from 1.81 to 13.67 and cooking loss was from 12.80 to 24.59, and both indicators were significantly correlated (R = 0.456, p < 0.01) due to the fact that both indicators related to moisture. In addition, tenderness is closely related to the palatability of meat products and its experimental measurement is a complex and imprecise process, as tenderness can be affected by many factors such as breed, age, muscle fiber diameter, muscle chemical composition and the slaughtering process. Therefore, the results of WBSF in Table 2 present the largest standard deviation (SD = 6.02). Furthermore, the results showed that the freezing temperature could affect a* (R = -0.192, p < 0.05), b* (R = -0.325, p < 0.01) and drip loss (R = 0.212, p < 0.05). The reason may be due to that the meat drip contained heme, and the drip loss could further result in the color change of thawed meat, leading to the decrease in red degree (a*) and yellowness (b*). 3.2. Spectral features of frozen pork meat The spectra of fresh sample, frozen samples, and frost were compared in the visible and near-infrared bands and are shown in Fig. 4. The spectral reflectance of the fresh was lower than those of the frozen samples in the whole Vis-NIR range (400 - 2500 nm), particularly, distinct troughs appeared near 440 nm, 570 nm, 980 nm, and 1200 nm. These wavelengths were considered to be the absorption regions of liquid water with the O-H stretching and bending [27]. When the sample was frozen, changes took place in the spectra: the trough at 440 nm moved to a shorter wavelength (near 420 nm). In addition, the reflectance near 980 nm and 1200 nm increased. Furthermore, a distinct peak appeared at 1880 nm for the frozen sample, while there was almost a straight line for the spectrum of the fresh one. Undoubtedly, the phase change process during freezing influenced the spectrum absorption of water, thus altering the shape of the whole spectral curve. Yamatera et al.
[28]
reported the effect of phase change on the
structure of water molecule. When water is frozen, the molecule is bound to its neighbors by hydrogen bonds, forming a stable three-dimensional network. When water is near its boiling point, molecules of water may undergo translational and rotational motions so violently that it is difficult to maintain suitable positions and orientations 7
for hydrogen bonding. The molecule interacts only weakly with its neighboring molecules. The current experiment confirmed that the higher the temperature, the more the spectrum absorption by the water, and therefore the lower the reflectance. In addition, an important finding was that the spectrum at 1880 nm contained the information of phase change and molecular structure of water. Even though the spectrum of the frost on sample’s surface was similar to that of the frozen sample, their details were different. For example, the overall reflection value of the frost was higher than that of the frozen meat. In addition, the trough near 420 nm was very shallow and the reflection peak near 1880 nm was more prominent. Therefore, if the frost spectrum was mixed with those of the frozen samples, it could cause difficulties to spectral modeling. 3.3. Influence of freezing temperatures on meat spectrum By ranking the spectra of the frozen samples based on the freezing temperatures, an interesting phenomenon was found. As shown in Fig. 5, there existed some order in the spectral patterns. The spectral reflectance near 1100 nm would increase with the lowering of the freezing temperatures, and the correlation coefficient of the reflectance with the freezing temperature was -0.832 (p < 0.01). By reviewing the experimental procedures, the meat samples were frozen at different temperatures (-120 to -20 °C) and then placed in a -20 °C refrigerator for 3 days to ensure that the surface temperature of every sample was -20 °C, and finally their spectra were scanned by the HSI system. This meant that the freezing temperature caused the spectral differences. However these spectral differences were not affected by the freezing methods used and the frozen storage temperatures employed. This phenomenon could be explained as follows: during the freezing process, different freezing rates induced various sizes of the ice crystals in the frozen samples, which affected the absorption of near-infrared spectrum, thus producing different corresponding spectral reflectance values. On the other hand, during frozen storage, the sizes and distribution of the ice crystals formed in the frozen samples would normally remain unchanged, therefore the spectral differences were not affected by the frozen storage temperatures. 3.4. Comparing model accuracy for diverse quality indicators In the current study, the entire region of the frozen sample was selected as ROI to extract the average spectrum. Based on the original and full-band spectrum (400 – 2500 nm), a preliminary model for each quality indicator was established. Table 4 shows the comparison of the model performances, indicating that the models for L* and cooking loss were the best, followed by those for drip loss and b*, and the model performances for pH and WBSF were the worst. 8
The standard deviation of the pH measurements was small (SD = 0.07), which was a disadvantage to develop robust model. In addition, the complexity in WBSF measurements could be the main reason for the poor performance of the corresponding model. Although both drip loss and cooking loss were related to moisture, the model prediction accuracy for drip loss was lower than that for cooking loss. The reason may be as follows: in the cooking process, all the cells of meat were damaged, the excess free water was squeezed out and the solid matter was heated to become edible, and therefore the prediction for cooking loss was closely related to the evaluation of the total moisture content. Cooking loss has been reported not to differ significantly between fresh and frozen meat samples, as well as for samples frozen and thawed at different rates [29]. On the other hand, the drip loss was just a part of the moisture in the meat sample, which was affected by the freezing rate, thawing rate and the damage degree of the microstructure of the muscle cells [30-31]. These differences contributed to the different performance of the models for the drip loss and the cooking loss. The correlation coefficient ( R p2 ) of the preliminary models for L*, a*, and b* was 0.734, 0.304, and 0.653, respectively, which indicated that it was possible to use HSI to predict the meat color after thawing. Based on the data in Table 4, it could be seen that the performances of the preliminary models for the quality indicators of the frozen samples were not as good as that for the fresh one [17], therefore further studies were needed to improve the prediction accuracy. 3.5. Different methods to select ROI Before modeling, an essential step was to extract spectrum from the ROIs as the average spectrum of the sample, which affected the performances of the models. In the current study, two methods for selecting ROI were compared: The first one was to select all pixels of the sample as the ROI by just removing the background, which is a method commonly used in many experiments, and the second one was besides removing the background, the frost and shadow areas were also excluded, and the remaining pixels were selected as the ROI. As discussed previously, some differences existed in the spectrum of the frozen meat and the frost, while the mixing of different spectrum would affect the performance of the models [32]. As shown in Table 4, based on the ROI selected by the second method, the accuracy of model prediction was improved. In particular, the improvement for the models for L* and cooking loss was most remarkable.
9
3.6. Modeling based on different spectral bands The spectra normally contain redundant and repetitive data. Therefore more wavebands do not necessarily mean better model performance. As shown in Table 5, the accuracy of the model for cooking loss based on the NIR spectrum was higher than that based on the full-band spectrum. For L*, the model based on the visible band spectrum was superior to that based on the full-band spectrum. This could be due to that the hydrogen bond absorption bands (1000 nm, 1200 nm) were located in the near-infrared range, while the visible range contained more color information. On the other hand, if the modeling was just based on three RGB wavelengths (700 nm, 510 nm, and 440 nm), the result was very unsatisfactory, as the ( R p2 ) of L* model based on RGB data was only 0.414, which was lower than the other three models. The result revealed that either Vis or NIR spectrum contains more useful information than that in RGB image. 3.7. Effects of pretreatment on spectral modeling Temperature fluctuations, surface roughness and the light-scattering could lower the performance of the models established. This problem could be reduced or eliminated by spectral preprocessing. In the current study, seven pretreatment methods were examined and the results are shown in Table 6. The results in Table 6 illustrates that not all of the pretreatment methods were effective in improving the accuracy of the models. The effect of Savitzky-Golay (S.Glay) and multiple scatter correction (MSC)
[33]
were the most significant, however, the
combined use of S.Glay and MSC would reduce the model accuracy. Therefore, it was particularly crucial to select the appropriate pretreatment method before modeling. Compared with the original spectra, MSC pretreatment made the spectra smoother and more in order, but their spectral patterns showed little difference. Therefore, ROI selection, waveband selection, and spectral pretreatment could effectively increase the performance of HSI models for rapid determination of frozen pork meat quality. The final model performances were not the same based on different quality indicators. The R p2 coefficient for L*, cooking loss, b*, drip loss, and a* was 0.907, 0.845, 0.814, 0.762, and 0.716, respectively; and their corresponding RMSEP was 0.900, 0.784, 0.643, 0.962 and 0.581, respectively. However, the prediction for the pH and WBSF suffered from low accuracy as 2
their R p coefficient was 0.257 and 0.213, and their RMSEP was 0.061 and 4.086, respectively. Although published studies [17, 34] confirmed HSI had a good performance in assessing pH of fresh meat, such good results did not appear on the frozen meat, which may be due to the difficulty in determining the pH value based on hydrogen ion concentration. When the meat is frozen, the hydrogen ions are bound in the 10
three-dimensional structure of the water molecules
[28]
, therefore it is difficult to detect the hydrogen ion
concentration of the meat in frozen state. Hence the prediction of pH for frozen meat was worse than that in fresh meat. The tenderness of the meat was affected by many factors, such as age, breed, sex, feeding situation, especially the complex biochemical reactions in the post-mortem ageing. In particular, these complex factors caused the large variations among subsamples on a single meat chop many previous works
[35-36]
[35]
in the measurement of WBSF. In fact,
demonstrated that the ability of spectroscopy technique was limited in prediction of
pork tenderness. While Douglas et al [34] showed spectroscopy combined with computer vision could improve the accuracy, future work is needed to confirm the approach for frozen meats. 3.8. Visualization of quality indicators The spectrum of each pixel was substituted into the regression models, the physical and chemical values of each point could then be predicted, as shown in Fig. 6. These distribution maps of the quality indicators can be very useful for evaluating the frozen pork meat qualities in details. By comparing the visualization maps, the image for cooking loss, L* and, b* value were in fine quality due to their low prediction errors. However interference fringes appeared in the visualization of drip loss, resulting in more blurry visualization than other three images. This blemish was more remarkable for lower accuracy indicators. 4. Conclusions The spectral characteristics of frozen pork meat were studied. The reflectance at 1100 nm had a very close relationship with the freezing temperature (R = -0.832, p < 0.01), this discovery was reported for the first time in the current study. Proper selection of the ROI pixels could withstand the interference of frost and extract effective information. PLSR was applied to establish spectral models and the results showed that choosing appropriate spectral bands and pretreatment methods could be crucial in developing robust models. Visible wavebands were suitable for predicting the color-related indicators, while NIR waveband had more advantages on the prediction of moisture-related indicators. The performances of the models established were varied based on different quality 2
indicators. The R p for L*, cooking loss, b*, drip loss and a* was 0.907, 0.845, 0.814, 0.762, and 0.716, 2
respectively, however R p for pH and WBSF were low, indicating the need for further studies. The current study confirmed that using HSI, it was possible to predict the quality of meats in frozen state directly without thawing, therefore the results in the study should have significantly expanded the applications of hyperspectral imaging 11
technology in the industry. Acknowledgements The authors gratefully acknowledge the Guangdong Province Government (China) for its support through the program “Leading Talent of Guangdong Province (Da-Wen Sun)”. This research was also supported by the National Key Technologies R&D Program (2014BAD08B09) and the International S&T Cooperation Projects of Guangdong Province (2013B051000010). References
[1]
ElMasry, G., Barbin, D. F., Sun, D. W., & Allen, P. (2012). Meat Quality Evaluation by Hyperspectral Imaging Technique: An Overview. Critical Reviews in Food Science and Nutrition, 52(8), 689-711.
[2]
Burkitt, L., Bunge, J., &Jargon, J. (2014). More woes for Yum and McDonald’s in China. Wall Street Journal Eastern Edition, 264(18), B3-B3.1.
[3]
Mortensen, M., Andersen, H. J., Engelsen, S. B., & Bertram, H. C. (2006). Effect of freezing temperature, thawing and cooking rate on water distribution in two pork qualities. Meat Science, 72(1), 34-42.
[4]
Muela, E., Sanudo, C., Campo, M. M., Medel, I., & Beltran, J. A. (2012). Effect of freezing method and frozen storage duration on lamb sensory quality. Meat Science, 90(1), 209-215.
[5]
Hansen, E., Trinderup, R. A., Hviid, M., Darre, M., & Skibsted, L. H. (2003). Thaw drip loss and protein characterization of drip from air-frozen, cryogen-frozen, and pressure-shift-frozen pork longissimus dorsi in relation to ice crystal size. European Food Research and Technology, 218(1), 2-6.
[6]
Oehlenschlager, J., & Mierke-Klemeyer, S. (2003). Changes of thaw-drip loss and cooking loss of Baltic cod (Gadus morhua) during long term storage under different frozen conditions. Deutsche Lebensmittel-Rundschau, 99(11), 435-438.
[7]
Huang, L., Xiong, Y. L. L., Kong, B. H., Huang, X. G., & Li, J. (2013). Influence of storage temperature and duration on lipid and protein oxidation and flavour changes in frozen pork dumpling filler. Meat Science, 95(2), 295-301.
[8]
Gambuteanu, C., Patrascu, L., & Alexe, P. (2014). Effect of freezing-thawing process on some quality aspects of pork Longissimus dorsi muscle. Romanian Biotechnological Letters, 19(1), 8916-8924.
[9]
Zhuang, H., & Savage, E. M. (2013). Comparison of cook loss, shear force, and sensory descriptive profiles of boneless skinless white meat cooked from a frozen or thawed state. Poultry Science, 92(11), 3003-3009.
[10]
Honikel, K. O. (1998). Reference methods for the assessment of physical characteristics of meat. Meat Science, 49(4), 447-457.
[11]
Watanabe A., Daly C. C., Devine C. E., (1996). The effects of the ultimate pH of meat on tenderness changes during ageing. Meat Science. 42(1):67-78.
[12]
Holmer, S. F., McKeith, R. O., Boler, D. D., Dilger, A. C., Eggert, J. M., Petry, D. B., McKeith, F. K., Jones, K. L., & Killefer, J. (2009). The effect of pH on shelf-life of pork during aging and simulated retail display. Meat Science, 82(1), 86-93.
[13]
Hardy, D., Kraeger, S., Dufour, S., & Cady, P. (1977). Rapid detection of microbial contamination in frozen vegetables by automated impedance measurements. Applied and environmental microbiology, 34(1), 14-17. 12
[14]
Koch, W. H., Payne, W. L., Wentz, B. A., & Cebula, T. A. (1993). Rapid polymerase chain reaction method for detection of Vibrio cholerae in foods. Applied and Environmental Microbiology, 59(2), 556-560.
[15]
Elmasry, G., Kamruzzaman, M., Sun, D. W., & Allen, P. (2012). Principles and applications of hyperspectral imaging in quality evaluation of agro-food products: a review. Critical Reviews in Food Science and Nutrition, 52(11), 999-1023.
[16]
Huang, H., Liu, L., Ngadi, M. O., & Gariepy, C. (2014). Rapid and non-invasive quantification of intramuscular fat content of intact pork cuts. Talanta, 119, 385-395.
[17]
ElMasry, G., Sun, D.-W., & Allen, P. (2012). Near-infrared hyperspectral imaging for predicting colour, pH and tenderness of fresh beef. Journal of Food Engineering, 110(1), 127-140.
[18]
li, Y., Shan, J., Peng Y., & Gao X. (2011). Nondestructive assessment of beef-marbling grade using hyperspectral imaging technology. Proceedings 2011 International Conference on New Technology of Agricultural Engineering (ICAE 2011), 779-783.
[19]
Barbin, D. F., ElMasry, G., Sun, D.-W., Allen, P., & Morsy, N. (2013). Non-destructive assessment of microbial contamination in porcine meat using NIR hyperspectral imaging. Innovative Food Science & Emerging Technologies, 17, 180-191.
[20]
Park, B., Windham, W. R., Lawrence, K. C., & Smith, D. (2007). Contaminant classification of poultry hyperspectral imagery using a spectral angle mapper algorithm. Biosystems Engineering, 96(3), 323–333.
[21]
Kamruzzaman, M., ElMasry, G., Sun, D. W., & Allen, P. (2012). Prediction of some quality attributes of lamb meat using near-infrared hyperspectral imaging and multivariate analysis. Analytica Chimica Acta, 714, 57-67.
[22]
Barbin, D. F., Sun, D.-W., & Su, C. (2013). NIR hyperspectral imaging as non-destructive evaluation tool for the recognition of fresh and frozen–thawed porcine longissimus dorsi muscles. Innovative Food Science & Emerging Technologies, 18, 226-236.
[23]
Cheng, J. H., Qu, J. H., Sun, D.-W., & Zeng, X. A. (2014). Visible/near-infrared hyperspectral imaging prediction of textural firmness of grass carp (Ctenopharyngodon idella) as affected by frozen storage. Food Research International, 56, 190-198.
[24]
Barbin, D. F., ElMasry, G., Sun, D.-W., & Allen, P. (2012). Near-infrared hyperspectral imaging for grading and classification of pork. Meat Science, 90(1), 259-268.
[25]
Kamruzzaman, M., ElMasry, G., Sun, D.-W., & Allen, P. (2011). Application of NIR hyperspectral imaging for discrimination of lamb muscles. Journal of Food Engineering, 104(3), 332-340.
[26]
Barbin, D. F., ElMasry, G., Sun, D.-W., & Allen, P. (2012). Predicting quality and sensory attributes of pork using near-infrared hyperspectral imaging. Analytica Chimica Acta, 719, 30-42.
[27]
Büning-Pfaue, H. (2003). Analysis of water in food by near infrared spectroscopy. Food Chemistry, 82(1), 107-115.
[28]
Yamatera, H., Fitzpatrick, B., & Gordon, G. (1964). Near infrared spectra of water and aqueous solutions. Journal of Molecular Spectroscopy, 14(1), 268-278.
[29]
Leygonie, C., Britz, T. J., & Hoffman, L. C. (2012). Impact of freezing and thawing on the quality of meat: Review. Meat Science, 91(2), 93-98.
[30]
Ngapo, T. M., Babare, I. H., Reynolds, J., & Mawson, R. F. (1999). Freezing and thawing rate effects on drip loss from samples of pork. Meat Science, 53(3), 149-158.
[31]
Ngapo, T. M., Babare, I. H., Reynolds, J., & Mawson, R. F. (1999). Freezing rate and frozen storage effects on the ultrastructure of samples of pork. Meat Science, 53(3), 159-168.
[32]
Roper, T., & Andrews, M. (2013). Shadow modelling and correction techniques in hyperspectral imaging. Electronics Letters, 49(7), 458-459.
13
[33]
Liu, X. M., & Liu, J. S. (2013). Measurement of soil properties using visible and short wave-near infrared spectroscopy and multivariate calibration. Measurement, 46(10), 3808-3814.
[34]
Barbin, D. F., Valous, N. A., & Sun, D. W. (2013). Tenderness prediction in porcine longissimus dorsi muscles using instrumental measurements along with NIR hyperspectral and computer vision imagery. Innovative Food Science and Emerging Technologies, 20, 335-342.
[35]
Chan, D. E., Walker, P. N., & Mills, E. W. (2002). Prediction of pork quality characteristics using visible and near-infrared spectroscopy. Transactions of the ASABE, 45(5), 1519–1527
[36]
Geesink, G. H., Schreutelkamp, F. H., Frankhuizen, R., Vedder, H. W., Faber, N. M., Kranen, R. W., et al. (2003). Prediction of pork quality attributes from near infrared reflectance spectra. Meat Science, 65, 661–668.
Table 1. Freezing parameters for pork samples Freezing technique
Freezing temperature
Number of samples
Equipment
Liquid nitrogen
-60 °C; -80 °C; -100 °C; -120 °C
30
DJL-SLX545, DeJianLi Co., Shenzhen, China
Immersion freezing*
-20 °C; -30 °C; -40 °C;
30
CTE-SE7510-05F, China-Scicooling Co., Beijing, China
Air-blast freezing
-20 °C; -40 °C; -60 °C;
30
CTE-SE7510-05F, China-Scicooling Co., Beijing, China
Domestic refrigerator
-20 °C; -25 °C; -30 °C;
30
BL/BD-719H, Haier Ltd. Qingdao, China
*Immersion freezing: the freezing medium was composed of ethanol and water in 1: 1 volume ratio. Pork samples were vacuum-packed before immersion freezing.
Table 2. Quality indicators measured by traditional methods for the frozen pork meat samples
Min Max Mean SD
pH
L*
a*
b*
Drip loss (%)
Cooking loss (%)
WBSF (N)
5.49 5.85 5.61 0.07
43.58 66.03 51.36 5.14
0.84 6.98 3.99 1.21
1.66 12.59 5.62 2.25
1.81 13.67 5.35 2.01
12.80 24.59 18.19 2.19
15.94 38.77 26.37 6.02
Min: minimum value; Max: maximum value; Mean: mean value; SD: standard deviation
14
Table 3. Correlation analysis on quality indicators of the frozen pork meat samples
pH L* a* b* Drip loss Cooking loss WBSF Temperature
Cooking loss
pH
L*
a*
b*
Drip loss
1
-0.351** 1
-0.062 -0.400** 1
-0.242* -0.022 0.571** 1
-0.250** 0.084 0.309** 0.533** 1
-0.202* -0.071 -0.043 -0.155 0.456** 1
-0.178
0.183
-0.192*
-0.325**
0.212*
-0.048
WBSF 0.174 0.049 -0.063 -0.297** -0.068 -0.080 1 0.159
* at a significant level (p
PII: DOI: Reference:
S0039-9140(15)00109-5 http://dx.doi.org/10.1016/j.talanta.2015.02.027 TAL15400
To appear in: Talanta Received date: 29 October 2014 Revised date: 6 February 2015 Accepted date: 17 February 2015 Cite this article as: Anguo Xie, Da-Wen Sun, Zhongyue Xu and Zhiwei Zhu, Rapid detection of frozen pork quality without thawing by Vis-NIR hyperspectral imaging technique, Talanta, http://dx.doi.org/10.1016/j.talanta.2015.02.027 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting galley proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
Rapid detection of frozen pork quality without thawing by Vis-NIR hyperspectral imaging technique Anguo Xiea, Da-Wen Suna,b, Zhongyue Xua, Zhiwei Zhua
a
College of Light Industry and Food Sciences, South China University of Technology, Guangzhou, 510641, China
b
Food Refrigeration and Computerised Food Technology, University College Dublin, National University of Ireland, Agriculture and Food Science Centre, Belfield, Dublin 4, Ireland
Abstract Quality determination of frozen food is a time-consuming and laborious work as it normally takes a long time to thaw the frozen samples before measurements can be carried out. In this research, a rapid and non-destructive determination technique for frozen pork quality was tested with a hyperspectral imaging (HSI) system. In this study, 120 pieces of pork meat were frozen by four kinds of methods with various freezing temperatures from -20 to -120 °C. The hyperspectral images of the samples were acquired at the frozen state. Quality indicators including drip loss, pH value, color, cooking loss and Warner-Bratzler shear force (WBSF) of the samples were measured after thawing. The spectral characteristics of the frozen meat samples were studied and it was revealed that the reflectance at 1100 nm had a close relationship with the freezing temperature (R = -0.832, p < 0.01). Partial least squares regression (PLSR) was applied to establish the spectral models, and the models were then optimized. Results showed that the improved region of interest (ROI) method could be used to extract effective spectral information to withstand the interference of freezing, and choosing appropriate spectral bands and spectral pretreatment techniques were crucial to develop robust mathematical model. The performances of the models established were diverse based on different quality indicators. The coefficients of determination for prediction 2
( R p ) for L*, cooking loss, b*, drip loss and a* were 0.907, 0.845, 0.814, 0.762, and 0.716, respectively. However 2
there were low correlations ( R p ) for pH and WBSF measurements. The current study indicated that HSI had the potential for non-destructive determination of frozen meat quality without thawing.
-1-7167342, Fax: +353-1-7167493, E-mail: [email protected],
Website: www.ucd.ie/refrig; www.ucd.ie/sun.
Keywords: Hyperspectral imaging; frozen food; freezing; NIR spectroscopy; drip loss; meat color
1. Introduction Pork is one of the most valuable meat products and is the primary choice for people to obtain protein. However pork meat is perishable [1], therefore freezing is a common technique to maintain its quality. When meats are in frozen state, it is not easy to determine their quality. Therefore the frozen meat market is mixed with superior and inferior products. It was reported that some food companies froze expired or inferior meats and sold them [2]. In addition, different freezing techniques can also have influences on the quality of frozen meats [3-4]. For example, slow freezing can promote the growth of large ice crystals, which could damage cell structure, upon thawing, it can cause more loss of meat juices
[5-6]
and color fading or darkening. Furthermore, temperature
fluctuations during refrigeration storage can also deteriorate the quality of the frozen foods, causing protein denaturation and fat oxidation [7], and spoilage [8-9]. Therefore, it is essential to ascertain the quality of frozen food products. Many indicators could be used to characterize the quality of frozen pork meats including drip loss after thawing, cooking loss, pH, color, and tenderness (Warner-Bratzler shear force). Among them, drip loss and cooking loss are two important indexes that have direct impact on the economic benefits of a company. Traditional methods are used to measure these quality indicators. Drip loss and cooking loss are measured by weighing the thawed or cooked sample and then comparing the weight with the initial weight of the frozen samples. In measuring drip loss, samples are usually thawed for 12 or 24 hours
[10]
. The pH value can be
determined by a pH meter, and its changes with storage time could indicate the freshness of the samples
[11-12]
.
Pork color can be measured by a colorimeter, which impacts on sensory evaluation and consumer desires. Tenderness reflects the palatability of the products, and can be determined by Warner-Bratzler shear force. All these measuring techniques are time-consuming and in particular, cannot meet the requirement of on-line measurements. Although some rapid detection studies have been carried out, for example Hardy et al. [13] detected the microbial contamination in frozen vegetables by automated impedance measurements and Koch et al. [14] used rapid polymerase chain reaction method to detect vibrio cholerae in foods, these studies focused on only microbiological detection, especially, it still needed a long time (more than 5 h) for testing. Therefore rapid and
2
non-destructive detection technique for more quality indicators is needed. Hyperspectral imaging (HSI) is a novel non-destructive evaluation method, which has been widely investigated for applications in the agricultural and food industry
[15]
. HSI provides both spatial and spectral
information for each pixel in the image. Therefore HSI can not only capture the external attributes (size, color, shape, surface texture, etc.) as traditional imaging technology, but can also be able to identify the chemical composition
[16]
in food as spectroscopic technique. Many studies have been conducted on using HSI for
determining meat quality. Elmasry et al.
[17]
developed a near infrared (NIR) HSI (900–1700 nm) system for the
measurement of surface color, pH and tenderness of fresh beef, Li et al.
[18]
used 400-1100nm HSI to assess
beef-marbling grade and obtained a good result with R2 = 0.92. Similar techniques were also employed to detect microbial contamination of porcine meat
[19]
, chicken contamination
[20]
, lamb meat quality
[21]
and so on. In
particular, Barbin et al. [22] evaluated the fresh and frozen–thawed porcine longissimus dorsi muscles by NIR HSI. Similar study was also reported by Cheng et al.
[23]
for fish fillets. However in these studies, the frozen samples
were thawed and then their images were acquired at room temperature by HSI. To the best of our knowledge, no research is available on using HSI technique to acquire food images at frozen state without thawing, and directly evaluate the frozen food quality. Therefore, the current study paper aimed to study a rapid and non-destructive technique based on hyperspectral imaging for direct measurement of frozen pork meat quality. The study also aimed to extract the spectral features of the frozen samples and to evaluate the impact of ice and frost on spectral detection.
2. Materials and methods The main steps of this experiment are shown in Fig. 1, which included freezing pork meat samples, image acquisition, measurement of quality indicators, spectral analysis, modeling and optimizing, and visualization. The experimental procedures are detailed below. 2.1. Freezing pork samples Longissimus dorsi meat samples from ten pigs were obtained from a local market at 0.5 day post-mortem. Then the pork samples were cut into 120 pieces (4 cm x 5 cm x 10 cm) with mass of 200 ± 4 g. These pork samples were frozen by four kinds of techniques including cryogenic freezing using liquid nitrogen, immersion freezing, air-blast freezing and domestic freezing in a conventional refrigerator. The main parameters of the
3
freezing processes are presented in Table 1. T-type thermocouple was inserted into the center of the sample to record the temperature variation with a data logger (TC-08, Pico Technology, Cambridgeshire, UK). There are a total of 120 samples, and each group of experiments contained 5 samples. When the core temperature of all 5 samples of one group reached -20 °C, the samples were moved into a refrigerator (BL/BD-719H, Haier Ltd., Qingdao, China) to store for 3 days at -20 °C. 2.2. Hyperspectral imaging system Spectral images were acquired by a hyperspectral imaging system shown in Fig. 2, which included a visible (Vis) unit (400 - 1000 nm) and a NIR unit (900 - 2500 nm). The Vis hyperspectral imaging system consisted of a camera lens (OLE23, Schneider, Rueil, German), a spectrograph (Imspector V10E, Spectral Imaging Ltd., Oulu, Finland) for the spectral range of 400–1000 nm, a high performance 1004 × 1002 charge coupled device (CCD) camera (DL-604 M, Andor, Belfast, Ireland), two 150 W halogen lamps (2900-ER, Illumination Technologies Inc., New York, USA) forming the illumination unit. The NIR unit consisted of a C-mount lens, a 12-bit CCD camera (XEVA 992, XC 130 XenICs, Leuven, Belgium) with 320 × 300 pixels, a spectrograph (ImSpector N17E, Specim, Oulu, Finland) covering the spectral range of 900–2500 nm, and an illumination unit of two 500 W tungsten halogen lamps (Lowel V-light™, NY, USA). The common parts of this Vis-NIR hyperspectral imaging system were a translation stage operated by a steppermotor (IRCP0076-1COMB, Isuzu Optics Co., Taiwan, China), and a computer supported with a data acquisition software (Spectral Image software, Isuzu Optics Co., Taiwan, China). 2.3. Image acquisition and calibration The difference of the Vis-NIR HSI system used in the current study from common hyperspectral imaging systems used in most studies was that a mini refrigerator (FYL-YS-30L, Fuyilian Co., Beijing, China) was installed on the translation stage. The temperature of the refrigerator was set at -20 °C. In the process of acquiring image, frozen pork sample was put in the refrigerator to avoid thawing. Every sample was scanned twice, i.e., the Vis image was acquired at the first, and then the NIR image was obtained by switching from the Vis system to the NIR system. In order to reduce the noise of the images, the acquired images were calibrated to obtain the calibrated image (R) using the following equation:
R
R 0 Rd 100% Rw Rd
(1)
4
where R0 is the raw hyperspectral image; Rd is the dark reference image of 0% reflectance, which was acquired by completely closing the lens of the camera with the opaque cap; and Rw is the white reference image (99%) obtained by using a white Teflon calibration tile [24]. 2.4. Measurement of quality parameters In this experiment, five quality indicators including drip loss, pH value, color value, cooking loss, and tenderness (Warner-Bratzler shear force) were measured. After image acquisition, the mass of the frozen pork samples (m1) was immediately weighed by an electronic balance (AL204, Mettler Toledo, Shanghai, China). Then the samples were thawed at the temperature of 4 °C for 24 hours
[5]
, the mass of the thawed samples (m2) was
measured again and the drip loss was calculated according to the equation below: Drip loss (%) =
m2 m1 100% m1
(2)
The pH value of the thawed sample was measured using a hand-held pH meter (Model 205, Testo AG, Lenzkirch, Germany), and the meat color was measured by a Chroma meter (CR-400, Konica Minolta Optics, Tokyo, Japan) in L*, a*, and b*. The measurement of the cooking loss was based on the procedure described by Honikel et al.
[10]
, i.e., the
frozen-thawed sample was cooked in a water bath at 80 °C until the core temperature of the sample reached 75 °C. The cooked sample was chilled at 4 °C for 4 h and then mass (m3) was taken by the balance. The cooking loss was calculated according to the following equation: Cooking loss (%) =
m3 m1 100% m1
(3)
where m1 is the mass of the frozen sample. Warner–Bratzler shear force (WBSF) is a recognized measurement for the tenderness of meats. In this experiment, every cooked sample was cut into three meat cuboids with a l00 mm2 (10 mm x 10 mm) cross-section area with the muscle fibre direction parallel to the long dimension of at least 30 mm
[10]
. Then the meat cuboids
were sheared with an Instron universal testing machine (5944S3013, Instron, Boston, USA) at the right angle to the fibre direction, and the shear force (N) was obtained to represent the tenderness. 2.5. Image segmentation and spectra extraction
5
In order to obtain useful spectral information, two kinds of methods were used to extract the spectra: (1) the background was removed from the hyperspectral image by subtracting a low-reflectance band from a high-reflectance band and the image was segmented by thresholding. All pixels of the frozen meat sample were selected as the region of interest (ROI), and then the spectra in ROI were averaged as the spectra of the whole sample
[25-26]
; (2) it was an improved ROI selecting method, in which not only the background but also the frost
region and shadow region were removed by using ENVI 4.8 software (Exelis Visual Information Solutions, Boulder, USA) as shown in Figure 3. The operation procedure in the 2nd method was to first open the ROI Tool in the ENVI software, then select preliminary ROI and finally press down "Grow" function key. The pixels in pre-existing ROI would be divided according to the deviation of spectra. 2.6. Establishment of spectral model and image visualization As one of the most common regression methods in chemometrics, partial least squares regression (PLSR) was used in the current study for predicting the quality indicators of the frozen pork meats. Among the 120 samples, they were randomly sorted, and then 90 samples were used as the calibration set and the remaining 30 samples were used for validation. The PLSR models were verified by cross-validation to confirm latent variables (LVs). The coefficients of determination for prediction ( R p2 ) and the root mean square error of prediction (RMSEP) were used as the criteria to assess the performance of the models. These algorithms were written and implemented in Matlab 2010b (MathWorks, Natick, USA). As a hyperspectral image is a three-dimensional matrix, which includes two-dimensional pixel matrix and one-dimensional spectral data, the spectra of each pixel were used as the input to the prediction models, and the result of each pixel was quantified. Then image visualization and prediction map were developed [13]. 3. Results and discussion 3.1. Quality indicators The quality indicators of the frozen pork samples are shown in Table 2 and Table 3. As indicated in Table 2, the mean value of pH for 120 samples was 5.61 ± 0.07. Many studies [11-12] have confirmed that pH of pork meat is greatly affected by storage time. For living pig, the pH of muscle is about 7.0. After slaughter, rigor mortis occurs and pH value decreases rapidly, and then pH increases slowly due to the ageing and spoilage of meat. In the current study, as all pork samples were stored for the same period after slaughtering, the pH values of the samples were similar. In addition, the data in Table 2 show that freezing and thawing did not affect the pH of the meat. 6
Meat color is an important indicator that can impact on consumer desires, and L*, a* and b* are used to indicate the brightness, red degree, and yellowness of the meat samples, respectively. Generally speaking, meat color is determined by the age of an animal, the amount of exercise it gets, the protein and fat content, and the storage period. Especially oxidation of myoglobin is a major factor for red color. As shown in Table 2, the average value of L*, a* and b* was 51.36, 3.99 and 5.62, respectively, and there existed strong correlations among the three color indicators of the meat samples, i.e., L* was significantly correlated with a* (R = -0.4, p < 0.01), and a* was highly correlated with b* (R = 0.571, p < 0.01). Table 2 also shows that the drip loss was from 1.81 to 13.67 and cooking loss was from 12.80 to 24.59, and both indicators were significantly correlated (R = 0.456, p < 0.01) due to the fact that both indicators related to moisture. In addition, tenderness is closely related to the palatability of meat products and its experimental measurement is a complex and imprecise process, as tenderness can be affected by many factors such as breed, age, muscle fiber diameter, muscle chemical composition and the slaughtering process. Therefore, the results of WBSF in Table 2 present the largest standard deviation (SD = 6.02). Furthermore, the results showed that the freezing temperature could affect a* (R = -0.192, p < 0.05), b* (R = -0.325, p < 0.01) and drip loss (R = 0.212, p < 0.05). The reason may be due to that the meat drip contained heme, and the drip loss could further result in the color change of thawed meat, leading to the decrease in red degree (a*) and yellowness (b*). 3.2. Spectral features of frozen pork meat The spectra of fresh sample, frozen samples, and frost were compared in the visible and near-infrared bands and are shown in Fig. 4. The spectral reflectance of the fresh was lower than those of the frozen samples in the whole Vis-NIR range (400 - 2500 nm), particularly, distinct troughs appeared near 440 nm, 570 nm, 980 nm, and 1200 nm. These wavelengths were considered to be the absorption regions of liquid water with the O-H stretching and bending [27]. When the sample was frozen, changes took place in the spectra: the trough at 440 nm moved to a shorter wavelength (near 420 nm). In addition, the reflectance near 980 nm and 1200 nm increased. Furthermore, a distinct peak appeared at 1880 nm for the frozen sample, while there was almost a straight line for the spectrum of the fresh one. Undoubtedly, the phase change process during freezing influenced the spectrum absorption of water, thus altering the shape of the whole spectral curve. Yamatera et al.
[28]
reported the effect of phase change on the
structure of water molecule. When water is frozen, the molecule is bound to its neighbors by hydrogen bonds, forming a stable three-dimensional network. When water is near its boiling point, molecules of water may undergo translational and rotational motions so violently that it is difficult to maintain suitable positions and orientations 7
for hydrogen bonding. The molecule interacts only weakly with its neighboring molecules. The current experiment confirmed that the higher the temperature, the more the spectrum absorption by the water, and therefore the lower the reflectance. In addition, an important finding was that the spectrum at 1880 nm contained the information of phase change and molecular structure of water. Even though the spectrum of the frost on sample’s surface was similar to that of the frozen sample, their details were different. For example, the overall reflection value of the frost was higher than that of the frozen meat. In addition, the trough near 420 nm was very shallow and the reflection peak near 1880 nm was more prominent. Therefore, if the frost spectrum was mixed with those of the frozen samples, it could cause difficulties to spectral modeling. 3.3. Influence of freezing temperatures on meat spectrum By ranking the spectra of the frozen samples based on the freezing temperatures, an interesting phenomenon was found. As shown in Fig. 5, there existed some order in the spectral patterns. The spectral reflectance near 1100 nm would increase with the lowering of the freezing temperatures, and the correlation coefficient of the reflectance with the freezing temperature was -0.832 (p < 0.01). By reviewing the experimental procedures, the meat samples were frozen at different temperatures (-120 to -20 °C) and then placed in a -20 °C refrigerator for 3 days to ensure that the surface temperature of every sample was -20 °C, and finally their spectra were scanned by the HSI system. This meant that the freezing temperature caused the spectral differences. However these spectral differences were not affected by the freezing methods used and the frozen storage temperatures employed. This phenomenon could be explained as follows: during the freezing process, different freezing rates induced various sizes of the ice crystals in the frozen samples, which affected the absorption of near-infrared spectrum, thus producing different corresponding spectral reflectance values. On the other hand, during frozen storage, the sizes and distribution of the ice crystals formed in the frozen samples would normally remain unchanged, therefore the spectral differences were not affected by the frozen storage temperatures. 3.4. Comparing model accuracy for diverse quality indicators In the current study, the entire region of the frozen sample was selected as ROI to extract the average spectrum. Based on the original and full-band spectrum (400 – 2500 nm), a preliminary model for each quality indicator was established. Table 4 shows the comparison of the model performances, indicating that the models for L* and cooking loss were the best, followed by those for drip loss and b*, and the model performances for pH and WBSF were the worst. 8
The standard deviation of the pH measurements was small (SD = 0.07), which was a disadvantage to develop robust model. In addition, the complexity in WBSF measurements could be the main reason for the poor performance of the corresponding model. Although both drip loss and cooking loss were related to moisture, the model prediction accuracy for drip loss was lower than that for cooking loss. The reason may be as follows: in the cooking process, all the cells of meat were damaged, the excess free water was squeezed out and the solid matter was heated to become edible, and therefore the prediction for cooking loss was closely related to the evaluation of the total moisture content. Cooking loss has been reported not to differ significantly between fresh and frozen meat samples, as well as for samples frozen and thawed at different rates [29]. On the other hand, the drip loss was just a part of the moisture in the meat sample, which was affected by the freezing rate, thawing rate and the damage degree of the microstructure of the muscle cells [30-31]. These differences contributed to the different performance of the models for the drip loss and the cooking loss. The correlation coefficient ( R p2 ) of the preliminary models for L*, a*, and b* was 0.734, 0.304, and 0.653, respectively, which indicated that it was possible to use HSI to predict the meat color after thawing. Based on the data in Table 4, it could be seen that the performances of the preliminary models for the quality indicators of the frozen samples were not as good as that for the fresh one [17], therefore further studies were needed to improve the prediction accuracy. 3.5. Different methods to select ROI Before modeling, an essential step was to extract spectrum from the ROIs as the average spectrum of the sample, which affected the performances of the models. In the current study, two methods for selecting ROI were compared: The first one was to select all pixels of the sample as the ROI by just removing the background, which is a method commonly used in many experiments, and the second one was besides removing the background, the frost and shadow areas were also excluded, and the remaining pixels were selected as the ROI. As discussed previously, some differences existed in the spectrum of the frozen meat and the frost, while the mixing of different spectrum would affect the performance of the models [32]. As shown in Table 4, based on the ROI selected by the second method, the accuracy of model prediction was improved. In particular, the improvement for the models for L* and cooking loss was most remarkable.
9
3.6. Modeling based on different spectral bands The spectra normally contain redundant and repetitive data. Therefore more wavebands do not necessarily mean better model performance. As shown in Table 5, the accuracy of the model for cooking loss based on the NIR spectrum was higher than that based on the full-band spectrum. For L*, the model based on the visible band spectrum was superior to that based on the full-band spectrum. This could be due to that the hydrogen bond absorption bands (1000 nm, 1200 nm) were located in the near-infrared range, while the visible range contained more color information. On the other hand, if the modeling was just based on three RGB wavelengths (700 nm, 510 nm, and 440 nm), the result was very unsatisfactory, as the ( R p2 ) of L* model based on RGB data was only 0.414, which was lower than the other three models. The result revealed that either Vis or NIR spectrum contains more useful information than that in RGB image. 3.7. Effects of pretreatment on spectral modeling Temperature fluctuations, surface roughness and the light-scattering could lower the performance of the models established. This problem could be reduced or eliminated by spectral preprocessing. In the current study, seven pretreatment methods were examined and the results are shown in Table 6. The results in Table 6 illustrates that not all of the pretreatment methods were effective in improving the accuracy of the models. The effect of Savitzky-Golay (S.Glay) and multiple scatter correction (MSC)
[33]
were the most significant, however, the
combined use of S.Glay and MSC would reduce the model accuracy. Therefore, it was particularly crucial to select the appropriate pretreatment method before modeling. Compared with the original spectra, MSC pretreatment made the spectra smoother and more in order, but their spectral patterns showed little difference. Therefore, ROI selection, waveband selection, and spectral pretreatment could effectively increase the performance of HSI models for rapid determination of frozen pork meat quality. The final model performances were not the same based on different quality indicators. The R p2 coefficient for L*, cooking loss, b*, drip loss, and a* was 0.907, 0.845, 0.814, 0.762, and 0.716, respectively; and their corresponding RMSEP was 0.900, 0.784, 0.643, 0.962 and 0.581, respectively. However, the prediction for the pH and WBSF suffered from low accuracy as 2
their R p coefficient was 0.257 and 0.213, and their RMSEP was 0.061 and 4.086, respectively. Although published studies [17, 34] confirmed HSI had a good performance in assessing pH of fresh meat, such good results did not appear on the frozen meat, which may be due to the difficulty in determining the pH value based on hydrogen ion concentration. When the meat is frozen, the hydrogen ions are bound in the 10
three-dimensional structure of the water molecules
[28]
, therefore it is difficult to detect the hydrogen ion
concentration of the meat in frozen state. Hence the prediction of pH for frozen meat was worse than that in fresh meat. The tenderness of the meat was affected by many factors, such as age, breed, sex, feeding situation, especially the complex biochemical reactions in the post-mortem ageing. In particular, these complex factors caused the large variations among subsamples on a single meat chop many previous works
[35-36]
[35]
in the measurement of WBSF. In fact,
demonstrated that the ability of spectroscopy technique was limited in prediction of
pork tenderness. While Douglas et al [34] showed spectroscopy combined with computer vision could improve the accuracy, future work is needed to confirm the approach for frozen meats. 3.8. Visualization of quality indicators The spectrum of each pixel was substituted into the regression models, the physical and chemical values of each point could then be predicted, as shown in Fig. 6. These distribution maps of the quality indicators can be very useful for evaluating the frozen pork meat qualities in details. By comparing the visualization maps, the image for cooking loss, L* and, b* value were in fine quality due to their low prediction errors. However interference fringes appeared in the visualization of drip loss, resulting in more blurry visualization than other three images. This blemish was more remarkable for lower accuracy indicators. 4. Conclusions The spectral characteristics of frozen pork meat were studied. The reflectance at 1100 nm had a very close relationship with the freezing temperature (R = -0.832, p < 0.01), this discovery was reported for the first time in the current study. Proper selection of the ROI pixels could withstand the interference of frost and extract effective information. PLSR was applied to establish spectral models and the results showed that choosing appropriate spectral bands and pretreatment methods could be crucial in developing robust models. Visible wavebands were suitable for predicting the color-related indicators, while NIR waveband had more advantages on the prediction of moisture-related indicators. The performances of the models established were varied based on different quality 2
indicators. The R p for L*, cooking loss, b*, drip loss and a* was 0.907, 0.845, 0.814, 0.762, and 0.716, 2
respectively, however R p for pH and WBSF were low, indicating the need for further studies. The current study confirmed that using HSI, it was possible to predict the quality of meats in frozen state directly without thawing, therefore the results in the study should have significantly expanded the applications of hyperspectral imaging 11
technology in the industry. Acknowledgements The authors gratefully acknowledge the Guangdong Province Government (China) for its support through the program “Leading Talent of Guangdong Province (Da-Wen Sun)”. This research was also supported by the National Key Technologies R&D Program (2014BAD08B09) and the International S&T Cooperation Projects of Guangdong Province (2013B051000010). References
[1]
ElMasry, G., Barbin, D. F., Sun, D. W., & Allen, P. (2012). Meat Quality Evaluation by Hyperspectral Imaging Technique: An Overview. Critical Reviews in Food Science and Nutrition, 52(8), 689-711.
[2]
Burkitt, L., Bunge, J., &Jargon, J. (2014). More woes for Yum and McDonald’s in China. Wall Street Journal Eastern Edition, 264(18), B3-B3.1.
[3]
Mortensen, M., Andersen, H. J., Engelsen, S. B., & Bertram, H. C. (2006). Effect of freezing temperature, thawing and cooking rate on water distribution in two pork qualities. Meat Science, 72(1), 34-42.
[4]
Muela, E., Sanudo, C., Campo, M. M., Medel, I., & Beltran, J. A. (2012). Effect of freezing method and frozen storage duration on lamb sensory quality. Meat Science, 90(1), 209-215.
[5]
Hansen, E., Trinderup, R. A., Hviid, M., Darre, M., & Skibsted, L. H. (2003). Thaw drip loss and protein characterization of drip from air-frozen, cryogen-frozen, and pressure-shift-frozen pork longissimus dorsi in relation to ice crystal size. European Food Research and Technology, 218(1), 2-6.
[6]
Oehlenschlager, J., & Mierke-Klemeyer, S. (2003). Changes of thaw-drip loss and cooking loss of Baltic cod (Gadus morhua) during long term storage under different frozen conditions. Deutsche Lebensmittel-Rundschau, 99(11), 435-438.
[7]
Huang, L., Xiong, Y. L. L., Kong, B. H., Huang, X. G., & Li, J. (2013). Influence of storage temperature and duration on lipid and protein oxidation and flavour changes in frozen pork dumpling filler. Meat Science, 95(2), 295-301.
[8]
Gambuteanu, C., Patrascu, L., & Alexe, P. (2014). Effect of freezing-thawing process on some quality aspects of pork Longissimus dorsi muscle. Romanian Biotechnological Letters, 19(1), 8916-8924.
[9]
Zhuang, H., & Savage, E. M. (2013). Comparison of cook loss, shear force, and sensory descriptive profiles of boneless skinless white meat cooked from a frozen or thawed state. Poultry Science, 92(11), 3003-3009.
[10]
Honikel, K. O. (1998). Reference methods for the assessment of physical characteristics of meat. Meat Science, 49(4), 447-457.
[11]
Watanabe A., Daly C. C., Devine C. E., (1996). The effects of the ultimate pH of meat on tenderness changes during ageing. Meat Science. 42(1):67-78.
[12]
Holmer, S. F., McKeith, R. O., Boler, D. D., Dilger, A. C., Eggert, J. M., Petry, D. B., McKeith, F. K., Jones, K. L., & Killefer, J. (2009). The effect of pH on shelf-life of pork during aging and simulated retail display. Meat Science, 82(1), 86-93.
[13]
Hardy, D., Kraeger, S., Dufour, S., & Cady, P. (1977). Rapid detection of microbial contamination in frozen vegetables by automated impedance measurements. Applied and environmental microbiology, 34(1), 14-17. 12
[14]
Koch, W. H., Payne, W. L., Wentz, B. A., & Cebula, T. A. (1993). Rapid polymerase chain reaction method for detection of Vibrio cholerae in foods. Applied and Environmental Microbiology, 59(2), 556-560.
[15]
Elmasry, G., Kamruzzaman, M., Sun, D. W., & Allen, P. (2012). Principles and applications of hyperspectral imaging in quality evaluation of agro-food products: a review. Critical Reviews in Food Science and Nutrition, 52(11), 999-1023.
[16]
Huang, H., Liu, L., Ngadi, M. O., & Gariepy, C. (2014). Rapid and non-invasive quantification of intramuscular fat content of intact pork cuts. Talanta, 119, 385-395.
[17]
ElMasry, G., Sun, D.-W., & Allen, P. (2012). Near-infrared hyperspectral imaging for predicting colour, pH and tenderness of fresh beef. Journal of Food Engineering, 110(1), 127-140.
[18]
li, Y., Shan, J., Peng Y., & Gao X. (2011). Nondestructive assessment of beef-marbling grade using hyperspectral imaging technology. Proceedings 2011 International Conference on New Technology of Agricultural Engineering (ICAE 2011), 779-783.
[19]
Barbin, D. F., ElMasry, G., Sun, D.-W., Allen, P., & Morsy, N. (2013). Non-destructive assessment of microbial contamination in porcine meat using NIR hyperspectral imaging. Innovative Food Science & Emerging Technologies, 17, 180-191.
[20]
Park, B., Windham, W. R., Lawrence, K. C., & Smith, D. (2007). Contaminant classification of poultry hyperspectral imagery using a spectral angle mapper algorithm. Biosystems Engineering, 96(3), 323–333.
[21]
Kamruzzaman, M., ElMasry, G., Sun, D. W., & Allen, P. (2012). Prediction of some quality attributes of lamb meat using near-infrared hyperspectral imaging and multivariate analysis. Analytica Chimica Acta, 714, 57-67.
[22]
Barbin, D. F., Sun, D.-W., & Su, C. (2013). NIR hyperspectral imaging as non-destructive evaluation tool for the recognition of fresh and frozen–thawed porcine longissimus dorsi muscles. Innovative Food Science & Emerging Technologies, 18, 226-236.
[23]
Cheng, J. H., Qu, J. H., Sun, D.-W., & Zeng, X. A. (2014). Visible/near-infrared hyperspectral imaging prediction of textural firmness of grass carp (Ctenopharyngodon idella) as affected by frozen storage. Food Research International, 56, 190-198.
[24]
Barbin, D. F., ElMasry, G., Sun, D.-W., & Allen, P. (2012). Near-infrared hyperspectral imaging for grading and classification of pork. Meat Science, 90(1), 259-268.
[25]
Kamruzzaman, M., ElMasry, G., Sun, D.-W., & Allen, P. (2011). Application of NIR hyperspectral imaging for discrimination of lamb muscles. Journal of Food Engineering, 104(3), 332-340.
[26]
Barbin, D. F., ElMasry, G., Sun, D.-W., & Allen, P. (2012). Predicting quality and sensory attributes of pork using near-infrared hyperspectral imaging. Analytica Chimica Acta, 719, 30-42.
[27]
Büning-Pfaue, H. (2003). Analysis of water in food by near infrared spectroscopy. Food Chemistry, 82(1), 107-115.
[28]
Yamatera, H., Fitzpatrick, B., & Gordon, G. (1964). Near infrared spectra of water and aqueous solutions. Journal of Molecular Spectroscopy, 14(1), 268-278.
[29]
Leygonie, C., Britz, T. J., & Hoffman, L. C. (2012). Impact of freezing and thawing on the quality of meat: Review. Meat Science, 91(2), 93-98.
[30]
Ngapo, T. M., Babare, I. H., Reynolds, J., & Mawson, R. F. (1999). Freezing and thawing rate effects on drip loss from samples of pork. Meat Science, 53(3), 149-158.
[31]
Ngapo, T. M., Babare, I. H., Reynolds, J., & Mawson, R. F. (1999). Freezing rate and frozen storage effects on the ultrastructure of samples of pork. Meat Science, 53(3), 159-168.
[32]
Roper, T., & Andrews, M. (2013). Shadow modelling and correction techniques in hyperspectral imaging. Electronics Letters, 49(7), 458-459.
13
[33]
Liu, X. M., & Liu, J. S. (2013). Measurement of soil properties using visible and short wave-near infrared spectroscopy and multivariate calibration. Measurement, 46(10), 3808-3814.
[34]
Barbin, D. F., Valous, N. A., & Sun, D. W. (2013). Tenderness prediction in porcine longissimus dorsi muscles using instrumental measurements along with NIR hyperspectral and computer vision imagery. Innovative Food Science and Emerging Technologies, 20, 335-342.
[35]
Chan, D. E., Walker, P. N., & Mills, E. W. (2002). Prediction of pork quality characteristics using visible and near-infrared spectroscopy. Transactions of the ASABE, 45(5), 1519–1527
[36]
Geesink, G. H., Schreutelkamp, F. H., Frankhuizen, R., Vedder, H. W., Faber, N. M., Kranen, R. W., et al. (2003). Prediction of pork quality attributes from near infrared reflectance spectra. Meat Science, 65, 661–668.
Table 1. Freezing parameters for pork samples Freezing technique
Freezing temperature
Number of samples
Equipment
Liquid nitrogen
-60 °C; -80 °C; -100 °C; -120 °C
30
DJL-SLX545, DeJianLi Co., Shenzhen, China
Immersion freezing*
-20 °C; -30 °C; -40 °C;
30
CTE-SE7510-05F, China-Scicooling Co., Beijing, China
Air-blast freezing
-20 °C; -40 °C; -60 °C;
30
CTE-SE7510-05F, China-Scicooling Co., Beijing, China
Domestic refrigerator
-20 °C; -25 °C; -30 °C;
30
BL/BD-719H, Haier Ltd. Qingdao, China
*Immersion freezing: the freezing medium was composed of ethanol and water in 1: 1 volume ratio. Pork samples were vacuum-packed before immersion freezing.
Table 2. Quality indicators measured by traditional methods for the frozen pork meat samples
Min Max Mean SD
pH
L*
a*
b*
Drip loss (%)
Cooking loss (%)
WBSF (N)
5.49 5.85 5.61 0.07
43.58 66.03 51.36 5.14
0.84 6.98 3.99 1.21
1.66 12.59 5.62 2.25
1.81 13.67 5.35 2.01
12.80 24.59 18.19 2.19
15.94 38.77 26.37 6.02
Min: minimum value; Max: maximum value; Mean: mean value; SD: standard deviation
14
Table 3. Correlation analysis on quality indicators of the frozen pork meat samples
pH L* a* b* Drip loss Cooking loss WBSF Temperature
Cooking loss
pH
L*
a*
b*
Drip loss
1
-0.351** 1
-0.062 -0.400** 1
-0.242* -0.022 0.571** 1
-0.250** 0.084 0.309** 0.533** 1
-0.202* -0.071 -0.043 -0.155 0.456** 1
-0.178
0.183
-0.192*
-0.325**
0.212*
-0.048
WBSF 0.174 0.049 -0.063 -0.297** -0.068 -0.080 1 0.159
* at a significant level (p
