Non-destructively sensing pork quality using near infrared multispectral imaging technique

Abstract

Near infrared multispectral imaging system (MSI) based on three wavebands—1280 nm, 1440 nm and 1660 nm—was developed for the non-destructive sensing of tenderness and water holding capacity (WHC) of pork. Multispectral images were acquired for pork samples, and the real tenderness (Warner-Bratzler Shear Force, WBSF) and WHC (cook loss, CL) of these samples were simultaneously determined using traditional destructive methods. The gray level co-occurrence matrix was used for the extraction of characteristic variables from multispectral images. Next, ant colony optimization combined with back propagation artificial neural network was used for modeling, which achieved good performance compared with the other two commonly used algorithms. The correlation coefficient and the root mean square error in the prediction set were achieved as follows: R_p = 0.8451 and RMSEP = 0.9087 for WBSF; R_p = 0.9116 and RMSEP = 1.5129 for CL. This work adequately demonstrates that the MSI technique has a high potential for non-destructive sensing of pork quality attributes combined with an appropriate algorithm, thus facilitating a simple and fast method of meat analysis.

---

1. Introduction

The quality of raw meat is usually influenced by a variety of factors which include the animal (breed, sex, age), environment factors (feeding, transporting and slaughtering condition), and processing conditions (storing time, temperature).¹ It is well known that all meat supplied to the markets must undergo quality control in order to guarantee consumer safety, and some consumers are willing to pay higher prices for meat products with an additional quality guarantee.²

There are many characteristics such as pH, color, texture, water holding capacity (WHC), tenderness, and freshness, for determining quality of meat.^2–4 Among them, tenderness and WHC are regarded as the most important quality parameters.⁵ Nowadays, tenderness and WHC of pork meat are commonly measured by destructive physicochemical means.

Tenderness correlates to many factors, such as the species and age of the animal at slaughter time. Usually, Warner-Bratzler Shear Force (WBSF) is seen as a reference index for evaluating pork tenderness. Determination of pork tenderness is performed by texture meters equipped with Warner-Bratzler, Kramer or compression devices.⁶

WHC is defined as the ability of muscle to retain water or resist water loss.⁷ In the past few years, WHC of pork has been measured using different procedures such as fluid loss, drip loss, thaw loss and cook loss (CL). However, these traditional methods are impractical because they are time consuming, laborious and destructive. Therefore, it is of great significance to explore a more rapid, efficient and non-destructive method for pork quality determination.

In the past decade, many researchers have successfully reported some non-destructive techniques used for the detection of pork quality. Among the non-destructive techniques, near infrared (NIR) spectroscopy is a widely used technique due to its rapidity, simplicity, and its ability to measure chemical properties or characteristics of food products.^8–10 In our previous studies, NIR spectroscopy was implemented to evaluate pork freshness and tenderness.^1,11 Nevertheless, NIR merely captures the single-point information of the sample, which is not sufficient to indicate the whole quality. Furthermore, change of pork quality is often accompanied with changes from external attributes (color, texture, etc.) and internal attributes (chemical compositions, tissue structure, etc.).¹² Therefore, it is key to seek a method to obtain outside and inside information. Recently, owing to its integration of traditional imaging and spectroscopy, hyperspectral imaging (HSI) technology has been widely used for food quality analysis.^13–17 However, the high price and large amount of data obtained slows down the detection speed, and thus limits it to a laboratory set-up which can not provide real-time information.¹⁸

Multispectral imaging (MSI) is an emerging platform technique with advantages of being rapid, chemical-free, non-destructive and so on compared with conventional analytical methods. In contrast to the NIR spectroscopy, it integrates traditional imaging and spectroscopy to attain both spatial and spectral information from objects. Although MSI only obtains a few images at a discrete spectral region by positioning a band-pass filter in front of a monochrome camera, the optimum waveband filters are found by HSI to develop an MSI system for practical use. Compared with the HSI technique, the MSI method is a low-cost system providing simple data suitable for real-time use. In recent years, MSI technology has been reported to have significant potential in food safety including food quality and safety of meat and meat products.^19–21 However, all of the above-mentioned within the range of 400–970 nm and most did not involve the quantitative analysis of pork quality.

The main aim of this work is to design a non-destructive, rapid, simple and low-cost method for pork quality analysis. Here, we developed a low-cost and simple MSI system from a complicated HSI system using data dimension reduction and selecting band-filters; we also systematically studied efficient non-linear algorithms for modeling. Accordingly, the specific objectives of the study include: selection of characteristic wavelengths using HSI; construction of MSI system and acquisition of multispectral images; pre-processing of multispectral images; extraction of characteristic parameters from the region of interest (ROI); the use of partial least squares (PLS), back propagation artificial neural network, and ant colony optimization combined with back propagation artificial neural network (ACO-BPANN) algorithms for modeling, as well as testing of the model using independent samples.

2. Materials and methods

2.1. Sample preparation

Samples (longissimus muscle) were purchased from fifteen pig carcasses in a local supermarket (Auchan, Xuefu road, Zhenjiang, China) and taken to the laboratory within 30 min. Pigs were slaughtered under commercial conditions (stunned electrically, exsanguinated, scalded, de-haired, eviscerated and split into sides), no other treatment at slaughter was carried out. Test samples were chopped into 103 pieces of 6 × 3 × 3 cm (length × width × thickness) on a sterile surface and the weight of each sample was about 80 ± 0.5 g. Before analysis, all samples were vacuum-packed with sealed plastic bags, labeled and stored in a refrigerator at 4 °C.

2.2. Characteristic wavelength selection

First of all, the HSI system, shown in Fig. 1a and developed by the Agricultural Product Processing and Storage Lab at Jiangsu University, was used for the acquisition of pork hyperspectral images. The system mainly consists of a high performance back-illuminated charge-couple device (CCD) camera (V10EB1610, Spectral Imaging Ltd., Finland) with a spatial resolution of 320 × 256 pixels; a line-scan spectrograph (ImSpector V10E2/3′′, Spectral Imaging Ltd., Finland) with a nominal spectral resolution of 5.0 nm; a 150 W quartz-halogen DC illuminator (Fiber-Lite PL900-A, Dolan-Jenner Industries Inc., USA); a linear motorized slide (Zolix SC30021A, Zolix. Corp., China); an enclosure; a data acquisition and pre-processing software; and a computer. The spectrograph collected spectral images in a wavelength range of 870–1770 nm, with a spectral interval of 3.5156 nm, which resulted in 256 spectral bands. Two fiber-optic light-guiding branches from the DC illuminator were mounted on the enclosure as light sources. The linear motorized slide was used to move the sample using a stepper motor controlled by the computer via a serial port so that both camera scanning and slide motion could be synchronized. A scanning rate was selected to achieve a square pixel. The whole imaging system was enclosed in a duralumin shield box (350 × 500 × 800 mm) to avoid interference from external light.

Fig. 1 NIR spectral imaging system.

The obtained hyperspectral data is a 3-D datacube including 256 images of the wavelength ranging from 870 nm to 1770 nm, while such huge data substantially increases the computational burden. Additionally, the neighboring band pictures from hyperspectral image data are highly correlated. The band-to-band correlation creates redundant information in the hyperspectral image data.²² Hence, it is necessary to extract optimum characteristic pictures from hyperspectral datacube that are related to pork quality. Principal component analysis (PCA) is one technique commonly used for dimensionality reduction intending to eliminate redundant bands and diminish computational burden. The front PC images, which express most information of original data, were found according to their variance contribution. Each PC image is a linear sum of the original images at individual wavelengths multiplied by the corresponding (spectral) weighting coefficients. Two or three bands with higher (local maximum) weighting coefficients from the optimum PC image are selected as the dominant bands. PCA was implemented in ENVI 4.5 (Research System, Inc., USA).

2.3. Multispectral imaging system and image acquisition

According to the dominant bands selected from the hypercube, the study designed an MSI system using the corresponding bandpass filters, which was developed by the Agricultural Product Processing and Storage Lab at Jiangsu University (see Fig. 1b). The MSI system mainly consists of a CCD camera (XS-1828XC117B, Xenics infrared solution, Belgium) with spatial resolution of 320 × 256 pixels, three 150 W quartz-halogen DC illuminators (Fiber-Lite PL900-A, Dolan-Jenner Industries Inc., USA), a rotating wheel of filters of characteristic bands (1280 ± 10 nm, 1440 ± 10 nm and 1660 ± 10 nm; Optical Insight, Inc., Santa Fe, NM), a computer, and image processing and analysis software (Matlab R2009b; The Math-works, Natick, MA, USA). The whole imaging system was enclosed in a duralumin shield box (350 × 500 × 800 mm) to avoid interference from external light.

Prior to image acquisition, the MSI system was preheated for 30 min. At the same time, the samples were taken out of the refrigerator and placed for 30 min at room temperature (25 ± 1 °C), and then they were placed on the conveying stage in the MSI system for multispectral data acquisition.

2.4. References measurement

WBSF is one reference method used to measure pork tenderness that has been reported in literature.^1,6,23 In this experiment, the WBSF measurement of pork was performed according to Chinese standard NY/T 1180-2006. After image acquisition, samples were immediately vacuum-packed in nylon/polyethylene bags, and cooked in a water bath at 80 °C until the internal temperature of the pork reached 70 °C. WBSF was measured by the TA-XT2i (Stable Micro Systems Limited Co., England) equipped with one Warner-Bratzler shear blade (cross-head speed of 1 mm s⁻¹). Then, the sample was sheared perpendicular to the muscle fibres. WBSF is determined according to the peak in the force deformation curve, and its unit, kilogram force, is abbreviated as kg f. In this work, each sample was measured five times, and the average of the five measurement results was used for further analysis.

Various methods have been used for the detection of pork WHC, including centrifuge force method, cooking loss, tray drip loss method and EZ (meat juice container procedure) drip loss method.⁹ In this experiment, the WHC measurement was performed according to the CL. For the determination of CL, samples were vacuum-packed in nylon/polyethylene bags and cooked by immersion at 80 °C until the internal temperature of pork reached 70 °C in the water bath. The pork samples were cooled to 25 °C. Next, the surface water of pork was wiped off with filter paper (Hangzhou Whatman-Xinhua Filter Paper Co. Ltd., Hangzhou, China). The difference of weight before and after cooking was used for CL calculation.

2.5. MSI images pre-processing

2.5.1 ROI segmentation. Some information from the original image is irrelevant to the analysis, such as its surroundings. To ensure that this irrelevant information will not interfere with the analysis, a pre-processing step is needed. Each individual spectral band image was extracted separately through defining an ROI. A quadrate ROI of size 50 × 50 pixels was selected by Matlab programs.

2.5.2 Extraction of texture feature variables. Gray level co-occurrence matrix (GLCM) has been widely used to extract image texture information.^4,24 Each element (i and j) in GLCM represents the probability that two pixels with the gray level i and j co-occur in the image separated by a distance along a given direction (0°, 45°, 90°, and 135°). Theoretically, a variety of GLCM could be constructed from the image with different values of direction and distance. In this study, four textural features including contrast, correlation, energy, and homogeneity were extracted by GLCM texture analysis. Generally, contrast is used to express the local variations present in the image. Correlation is a measure of image linearity among pixels and the lower the values, the less linear correlation. Energy that measures the textural uniformity of the image is the sum of squared elements in the GLCM. Finally, homogeneity usually measures the closeness of the distribution of elements in the GLCM to its diagonal. The above mentioned parameters were calculated at one distance (D = 1) for each pixel in the GLCM in each direction (0°, 45°, 90°, and 135°). The mean and standard deviation of each image were calculated. Then, the model prediction was conducted by Matlab programs using texture feature parameters (contrast, correlation, energy, and homogeneity under 4 directions, mean and standard deviation) from three characteristic pictures (totalling 54 variables for one sample) as the input variables, whereas the measured values of WBSF and WHC were the output variables. The steps involved in building prediction models are depicted in the flowchart shown in Fig. 2.

Fig. 2 Flowchart of predicting WBSF and CL in pork by MSI system.

2.6. Software

HSI data was acquired by SpectralCube (ImSpector, image, Auto Vision Inc., USA). Characteristic wavelength optimization was implemented in ENVI 4.5 (Research System, Inc., USA). MSI data acquisition software was compiled based on Microsoft VC++ platform, and all data algorithms were implemented in Matlab R2009b (Matworks Inc., Natick, MA, USA) in Windows 7.

3. Results and discussion

3.1. Calibration of models

All 103 samples were divided into two subsets. The first subset was called a calibration set used to build a model, while the other was called a prediction set used to test the robustness of the model. Selecting samples for modeling was done as follows: first, all samples were sorted according to their respective y-value (viz. the reference values of WBSF and CL); then, one sample out of every three was entered into the prediction set. Thus, the calibration set contained 69 samples, and the prediction set contained 34 samples. As shown in Table 1, the ranges of the reference values of WBSF and CL in the calibration set almost cover the range of the prediction set, and their standard deviations in the calibration and prediction sets have no significant differences. Therefore, their distributions of the samples are reasonable in the calibration and prediction sets.

Table 1 Reference measurement of WBSF and CL in the calibration and prediction sets

Quality parameters	Units	Subsets	Sample number	Range	Mean	Standard deviation
WBSF	kg f	Calibration set	69	4.4678–9.8537	6.8921	1.2392
		Prediction set	34	5.0626–10.1651	6.9992	1.0182
CL	%	Calibration set	69	21.33–29.69	26.94	1.87
		Prediction set	34	22.09–30.7	27.04	1.74

---

3.2. Results of wavelength optimization

In the spectral range 870–1700 nm, too much noise was found in the spectral regions below 900 nm and over 1700 nm, and thus, the spectral region 900–1700 nm was selected. In this study, PCA was used to reduce the hyperspectral data dimension. The first principal component (PC1) image is shown in Fig. 3, which is the first PC image obtained by PCA. It is found that the PC1 image is the best representation of the original sample, because the variance contribution rate is the highest, reaching 95.53%. Thus, the dominant bands are determined according to the PC1 image in this work, and four dominant bands (i.e., 976.759 nm, 1281.865 nm, 1436.192 nm and 1662.201 nm) with higher weight coefficients are selected by investigating all weighting coefficients. It is seen that the selected wavelengths for pork quality were closely related to the water and fat content of the samples. The absorption bands of 976.759 nm, 1281.865 nm, 1436.192 nm and 1662.201 nm are due to O–H bonds attributed to water content or C–H bonds related to fat content,⁴ and it is known that the main ingredients of meat are protein, fat, carbohydrates and water. Accordingly, the images at these 4 dominant bands can correspond to the quality of pork. In order to make the system simple, we chose three of the wavelengths to establish a MSI system. Thus, the multispectral band pass filters with bandwidths of 1280 ± 10 nm, 1440 ± 10 nm and 1660 ± 10 nm were selected.

Fig. 3 Dominant wavelengths selected by PCA.

3.3. Results of ACO-BPANN models

As mentioned above, a change in pork quality is often accompanied with changes in external attributes (color, texture, etc.) and internal attributes (chemical compositions, tissue structure, etc.). The three multispectral data-cube images can show the changes of external attributes of pork, and the spectral information can show the internal attribute changes. Thus, there exists an indirect correlation between the quality of pork meat and multispectral data. However, the variables found from characteristic images, may have a linear correlation. Moreover, the correlating information is redundant and unfavorable for establishing a simple model. In this work, the MSI technique combined with variable selection method (ant colony optimization (ACO)) was used to develop a BP-ANN model for prediction of WBSF and CL in pork.

ACO²⁵ is an optimization method that can be used for feature selection. It resembles the behavior of ant colonies in the search for the best path to food sources without the use of visual information, which employs the concept of cooperative pheromone accumulation, and optimizes models using a pre-defined number of variables, occupying a Monte Carlo approach to discard irrelevant variables.²⁶ It has an advantage over simulated annealing and genetic algorithm approaches for similar problems where the graph may change dynamically; the ant colony algorithm can be run continuously and adapt to changes in real-time.²⁷ Presently, ACO shows its high potential in discarding irrelevant information regions, and has been widely used for characteristic variable selection.²⁸ In this work, the required parameters for running ACO algorithms were set as follows according to experience based on substantial trials: the initial population was set to 80; the maximum number of iterations was set to 50; the maximum number of cycles was set to 20; the probability threshold of variable selection was set to 0.3; and the pheromone attenuation coefficient was set to 0.65.

For the HSI data processing, there are many algorithms reported in the literature, such as, PLS, support vector machines, ANN.^16,29 BP-ANN is a strong tool for capturing and revealing a complex relationship between inputs and outputs. BP-ANN is commonly used for feed-forward multilayer networks.^30,31 It consists of neurons arranged in layers (an input layer, one or more hidden layers and an output layer) with connections (weights) unidirectional from input to output.^31–33 The optimal model was determined by the lowest root mean square error of cross validation (RMSECV). Other parameters of the BP-ANN model were optimized by the minimal mean square error (MSE). Herein, the number of neurons in the hidden layer was set as 5, the learning rate factor and momentum factor were set as 0.1, the initial weight was set as 0.3 and the scale function was set as ‘tangential hyperbolic (tanh)’ function.

Fig. 4a presents the textural features variables selected by the ACO algorithm, which contains a total of 10 variables. It indicates that these 10 variables are highly correlated with WBSF of pork. Fig. 4b is the scatter plot between reference measurements of WBSF and MSI predicted results in the calibration and prediction sets. Here, the value of RMSECV is 0.921 kg f, and the correlation coefficient (R_c) is 0.9267 in the calibration set, when the performance of the BP-ANN model is evaluated by the samples in the prediction set, the value of RMSEP is 0.9087 kg f, and the correlation coefficient (R_p) is 0.8451. Fig. 4c presents the textural features variables selected by the ACO algorithm, which contains a total of 14 variables. It indicates that these 14 variables are highly correlated with CL of pork. Fig. 4d is the scatter plot between the reference measurements of CL and MSI predicted results in the calibration and prediction sets. Here, the value of RMSECV is 1.3391%, and the correlation coefficient (R_c) is 0.9533 in the calibration set; when the performance of BP-ANN model is evaluated by the samples in the prediction set, the value of RMSEP is 1.5129%, and the correlation coefficient (R_p) is 0.9116.

Fig. 4 Results of ACO-BPANN models: (a) variables selected by ACO algorithm for WBSF; (b) scatter plot between reference measurement of WBSF and ACO-BPANN predicted results; (c) variables selected by the ACO algorithm for CL; (d) the scatter plot between reference measurements of CL and ACO-BPANN predicted results.

3.4. Discussion of the results

In order to highlight the superiority of the ACO-BPANN algorithm, classical PLS and BP-ANN were studied systematically and comparatively. All the results are shown in Table 2. All the regression models show good performance, which further verified that it is feasible and reliable to analyze WBSF and CL of pork meat quantitatively with the selected wavelengths using the developed MSI system. Moreover, the results of BP-ANN models are much better than those of PLS models. In addition, in the case of the two parameters, the regression models made great progress after ACO variables selection. The results revealed that ACO-BPANN is extremely suitable for the determination of pork WBSF and CL by MSI technique.

Table 2 Calibration models for prediction of pork WBSF and CL using different algorithms

Quality parameters	Model	Variables	Prediction set, Rp
WBSF	PLS	54	0.7481
	BP-ANN	54	0.8253
	ACO-BPANN	10	0.8451
CL	PLS	54	0.7786
	BP-ANN	54	0.8876
	ACO-BPANN	14	0.9116

---

The reason for the results could be explained as follows. Firstly, it might be explained by the histological basis of pork. Muscle tissue is composed of myofibrils, and each myofibril consists of myosin heavy-chain and actin filaments. Differences in μ-calpain, m-calpain, and calpastatin activity may ultimately influence the tenderness and water-holding capacity of pork by impacting the rate of myofibril, adipose tissue and water.^34–36 Furthermore, after cooking, the adipose tissue cells rupture and the intramuscular fat redistributes, which also affects pork quality. The distribution of myofibrils, adipose tissue and water forms the texture which can be captured by MSI data, thus, the MSI images can indirectly reflect pork quality. Therefore, all the models achieved good performance, indicating that it is feasible and reliable to estimate WBSF and CL quantitatively using our developed MSI system.

Secondly, the BP-ANN model was compared with the PLS model. Compared to the linear regression tool (i.e., PLS), BP-ANN is a universal non-linear regression tool which has stronger robustness, self-learning and adaptation than the linear method. When faced with complex problems, the non-linear method might be more suitable for the solution of data prediction.^12,37 In contrast with the linear structure of PLS, the topological network architecture of BP-ANN may be more suitable for the analysis of complicated measurements.³⁸ In fact, the relationships between the WBSF/CL and the multispectral data were diagnosed by the approach of augmented partial residual plots (APARPs).³⁹ A quantitative numerical tool (run test) was employed to calculate the non-linearity based on the APARPs method; the |z|-value is 8.912 for WBSF, and 8.512 for CL, respectively; both of them exceed the critical value (|z| = 1.96). The results are shown in Table 3. Therefore, it can be concluded that there is a non-linear relationship between multispectral images and WBSF and CL, and the linear tools might not be able to provide a complete solution to so complicated a regression.

Table 3 Results of the run test used for detection of the linearity relationship between the MSI data and WBSF and CL based on the APARPs method

Quality parameters	n+	n−	u	μ	σ	|z|	Result
WBSF	52	51	7	52.5	5.05	8.91	Non-linear
CL	54	49	9	52.4	5.04	8.51	Non-linear

---

Thirdly, the ACO-BPANN model was compared with the BP-ANN model. The absorption bands at characteristic wavelengths of 1280 nm, 1440 nm and 1660 nm are due to O–H bonds related to water content.^5,14,40 Therefore, the texture feature variables, which were extracted from 3 MSI images, may have a linear correlation. Moreover, like ants finding food, the ACO algorithm is good for optimizing the variables. The variables optimized by ACO, which were closely related to pork quality, were used as the input for the model. Thus, compared with BP-ANN and PLS models, the ACO-BPANN model has the best result.

4. Conclusions

This work shows the potential of a near infrared MSI technique in the determination of WBSF and WHC in pork, which are two important indicators of pork quality. In developing prediction models, ACO-BPANN revealed its superiority in contrast to classical PLS and BP-ANN calibration methods. It can be concluded that the MSI system, with an efficient algorithm, could realise the application of emerging imaging technique from the laboratory to practical applications like real-time monitoring of meat quality.

Acknowledgements

This work has been financially supported by the National Natural Science Foundation of China (31371770). We are also grateful to many of our colleagues for stimulating discussion in this field.

References

1. J. Cai, Q. Chen, X. Wan and J. Zhao, Food Chem., 2011, 126, 1354–1360.
2. B. Leroy, S. Lambotte, O. Dotreppe, H. Lecocq, L. Istasse and A. Clinquart, Meat Sci., 2004, 66, 45–54.
3. F. Tao, Y. Peng, Y. Li, K. Chao and S. Dhakal, Meat Sci., 2012, 90, 851–857.
4. M. Kamruzzaman, G. ElMasry, D.-W. Sun and P. Allen, Anal. Chim. Acta, 2012, 714, 57–67.
5. M. Kamruzzaman, G. ElMasry, D.-W. Sun and P. Allen, Food Chem., 2013, 141, 389–396.
6. Y. L. Xiong, M. J. Gower, C. Li, C. A. Elmore, G. L. Cromwell and M. D. Lindemann, Meat Sci., 2006, 73, 600–604.
7. D. Tejerina, S. García-Torres and R. Cava, Livest. Sci., 2012, 148, 46–51.
8. D. K. Pedersen, S. Morel, H. J. Andersen and S. Balling Engelsen, Meat Sci., 2003, 65, 581–592.
9. M. Prevolnik, M. Čandek-Potokar and D. Škorjanc, J. Food Eng., 2010, 98, 347–352.
10. S. M. Fowler, H. Schmidt, R. van de Ven, P. Wynn and D. L. Hopkins, Meat Sci., 2014, 98, 652–656.
11. Q. Chen, J. Cai, X. Wan and J. Zhao, LWT--Food Sci. Technol., 2011, 44, 2053–2058.
12. Q. Chen, Z. Hui, J. Zhao and Q. Ouyang, LWT--Food Sci. Technol., 2014, 57, 502–507.
13. G. K. Naganathan, L. M. Grimes, J. Subbiah, C. R. Calkins, A. Samal and G. E. Meyer, Comput. Electron. Agr., 2008, 64, 225–233.
14. G. ElMasry, D.-W. Sun and P. Allen, J. Food Eng., 2012, 110, 127–140.
15. K. Cluff, G. Konda Naganathan, J. Subbiah, A. Samal and C. R. Calkins, Meat Sci., 2013, 95, 42–50.
16. D. Wu and D.-W. Sun, Talanta, 2013, 116, 266–276.
17. H. Li, Q. Chen, J. Zhao and M. Wu, LWT--Food Sci. Technol., 2015, 63, 268–274.
18. Q. Chen, C. Zhang, J. Zhao and Q. Ouyang, TrAC, Trends Anal. Chem., 2013, 52, 261–274.
19. X. Sun, K. J. Chen, K. R. Maddock-Carlin, V. L. Anderson, A. N. Lepper, C. A. Schwartz, W. L. Keller, B. R. Ilse, J. D. Magolski and E. P. Berg, Meat Sci., 2012, 92, 386–393.
20. F. Ma, J. Yao, T. Xie, C. Liu, W. Chen, C. Chen and L. Zheng, Food Res. Int., 2014, 62, 902–908.
21. E. Z. Panagou, O. Papadopoulou, J. M. Carstensen and G.-J. E. Nychas, Int. J. Food Microbiol., 2014, 174, 1–11.
22. L. Huang, J. Zhao, Q. Chen and Y. Zhang, Food Res. Int., 2013, 54, 821–828.
23. J. Wu, Y. Peng, Y. Li, W. Wang, J. Chen and S. Dhakal, J. Food Eng., 2012, 109, 267–273.
24. G. ElMasry, N. Wang, A. ElSayed and M. Ngadi, J. Food Eng., 2007, 81, 98–107.
25. M. Dorigo and L. M. Gambardella, IEEE Trans. Evol. Comput., 1997, 1, 53–66.
26. F. Allegrini and A. C. Olivieri, Anal. Chim. Acta, 2011, 699, 18–25.
27. V. S. D. R. Umarani and V. Selvi, International Journal of Computer Applications, 2010, 5, 1–6.
28. X. W. Huang, X. B. Zou, J. W. Zhao, J. Y. Shi, X. L. Zhang and M. Holmes, Food Chem., 2014, 164, 536–543.
29. Q. Ouyang, J. W. Zhao, W. X. Pan and Q. S. Chen, Food Chem., 2016, 190, 135–141.
30. F. Ahmed, Mol. Cancer, 2005, 4, 29.
31. L. Yu, S. Wang and K. K. Lai, Comput. Oper. Res., 2005, 32, 2523–2541.
32. K. Cai, J. Xia, L. Li and Z. Gui, Comput. Mater. Sci., 2005, 34, 166–172.
33. T.-L. Lee, Ocean Eng., 2004, 31, 225–238.
34. A. H. Karlsson, R. E. Klont and X. Fernandez, Livest. Prod. Sci., 1999, 60, 255–269.
35. J. L. Melody, S. M. Lonergan, L. J. Rowe, T. W. Huiatt, M. S. Mayes and E. Huff-Lonergan, J. Anim. Sci., 2004, 82, 1195–1205.
36. J. M. Hughes, S. K. Oiseth, P. P. Purslow and R. D. Warner, Meat Sci., 2014, 98, 520–532.
37. Q. Chen, Z. Guo, J. Zhao and Q. Ouyang, J. Pharm. Biomed. Anal., 2012, 60, 92–97.
38. H. Lin, Q. Chen, J. Zhao and P. Zhou, J. Pharm. Biomed. Anal., 2009, 50, 803–808.
39. V. Centner, O. E. de Noord and D. L. Massart, Anal. Chim. Acta, 1998, 376, 153–168.
40. A. Iqbal, D.-W. Sun and P. Allen, J. Food Eng., 2013, 117, 42–51.

---