Identification of Repeatedly Frozen Meat Based on Near-Infrared Spectroscopy Combined with Self-Organizing Competitive Neural Networks

Abstract

A rapid, effective method of identifying repeatedly frozen meat by near-infrared spectroscopy (NIRS) combined with a self-organizing competitive neural network (SCNN) model was established. A total of 180 samples were adopted, including hot, cold, frozen, and repeatedly frozen meats. We compared the treatment effects of four pretreatment methods on the spectrogram samples, namely, multiplicative scatter correction (MSC), standard normal variables (SNV), first-order differential and second-order differential. The second differential pretreatment exerted the optimum effect. A total of 120 pork samples were randomly selected and used to establish a calibration model, and the remaining 60 samples were used for prediction. SCNN analysis revealed that classification performance was the highest when the learning number was 250. The recognition ratio of the 60 prediction collection was 93.3%, in which the recognition ratio of the repeatedly frozen meat was 100%. Thus, combined NIRS and SCNN can rapidly and accurately detect repeatedly frozen meat without destruction.

Keywords:

• Repeatedly frozen meat
• Near-infrared spectroscopy
• Principal component analysis
• Self-organizing competitive neural network

INTRODUCTION

As the most popular meat resource in the world, pork requires intensive quality supervision. The quality of pork has important implications in the quality of daily life, nutrition, ingestion, and dietary safety. According to ordinary treatment methods, pork can be divided into hot meat (HM), cold meat (CM), and frozen meat (FM). HM is fresh meat slaughtered at night and distributed in the early morning. CM refers to fresh meat rapidly cooled after slaughter to remove heat and then refrigerated at 0 to 4°C. FM refers to pork initially placed below –25°C to freeze and then stored at –18°C. These three kinds of meat are usually suitable for human consumption.

However, some merchants illegally sell repeatedly FM. Repeated freezing reduces the water-holding capacity, nutritional value, and flavor of meat. Microbial fecundity and enzyme activity increases during thawing to refreezing. Thus, refrozen meat easily becomes unhealthy and may even produce carcinogenic substances. Existing researches have shown that there would be peroxide fat in the refrozen meat which may lead to cancer.^[1,2]

A significant quality difference exists between refrozen meat and FM, but these meat types cannot be distinguished with the naked human eye. Thus, meat quality determination must be enhanced. Traditional meat quality determination methods mainly include biochemistry and molecular biology methods, such as gel electrophoresis,^[3,4] enzyme-linked immunoassay,^[5,6] DNA probe,^[7,8] and chemical composition analysis.^[9] However, these aforementioned methods are time consuming and involve complicated operations and high costs.

Near-infrared (NIR) spectroscopy is an effective, rapid, non-destructive, and new analytical technique widely used in the rapid quantitative and qualitative analysis of agricultural products.^[10–12] In meat analysis, NIR spectroscopy is used to detect the main chemical compositions of meat and meat products, such as protein, fat, and moisture,^[13–17] as well as to identify meat (qualitatively varieties, years, genotypes, type of feeding manners, positions, and production sites.^[18–20] Many researchers have reported on the quantitative and qualitative analyses of meat and meat products. The identification research of fresh beef and frozen-thawed beef based on NIR spectroscopy combined with factorial discriminant analysis (FDA) and Soft Independent Modeling of Class Analogy (SIMCA) have been done by Downey and Beauchne.^[21] Zhang and Cheng^[22] have done the identification of fresh shrimp and frozen-thawed shrimp by Vis/NIR spectroscopy combined with discriminant partial least-squares (DPLS) and least squares-support vector machine (LS-SVM) methods. However, no study has reported on the identification of refrozen pork meat.

In this study, a technique was developed for the rapid and effective identification of various kinds of pork. Given that disreputable businesses produce and sell repeatedly frozen pork, the authors aimed to explore a possible method of identifying repeatedly frozen pork using NIR spectroscopy combined with principal component analysis (PCA) and a self-organizing competitive neural network (SCNN) model to classify pork.

MATERIALS AND METHODS

Materials and Sample Preparation

Fresh pork directly purchased from a supermarket was labeled HM. Fresh pork stored in the refrigerator at 4°C for 8 h was labeled CM. Fresh pork initially frozen in the refrigerator at –25°C for 12 h and then stored at –18°C for 8 h was labeled FM. FM was first thawed at room temperature for 8 h and then placed in the freezer at –25°C for 12 h was labeled as repeatedly FM or complex FM (CFM).

Spectrum Acquisition

Classified pork samples were cut into small pieces with a meat grinder, and their spectra were obtained by XDS Rapid Content Grating NIR spectroscopy (Denmark FOSS) ranging from 760 to 2500 nm with one dot recorded every 2 nm. We collected 45 spectra from the four kinds of pork samples. Each spectrum had 870 absorbance values. Measurement experiments were conducted under ambient conditions of 23 ± 1°C and relative humidity of 46%.

PCA

PCA is an important method of multivariate statistical analysis.^[23,24] Multiple indicators can be transformed to a few aggregated indicators that retain most of the information of the original indicators by reducing dimensions. According to the variance maximum principle, linear fittings are used for several independent variables included in the original spectrum data. New low-dimensional variables replace the original high-dimensional variables to acquire data with reduced dimensions.^[25,26] The main component is the dominant contributor to the amount of deviation. Thus, data of high-dimensional space can be reduced to a low dimension. The authors aimed to observe the data easily and reduce the amount of information loss at T > 80%.^[27] This method eliminates overlapping and redundant indices in a project with multiple indicators. The correlation and variation in the data determine the weight without subjective human judgment, thereby resulting in more objective evaluation.

Self-Organizing Neural Network

A self-organizing neural network is a free teaching-learning neural network model that adopts competitive learning rules.^[28–30] This neural network can learn and simulate unknown samples and make appropriate adjustments for its own network structure. As shown in Fig. 1, the network structure possesses an input layer and a competitive layer. The input layer accepts outside information and passes the input mode to the competitive layer, which acts as an observer. The competitive layer performs mode analysis and comparison to identify the rules for correct classification.

FIGURE 1 Schematic illustration of the structure of the self-organizing neural network.

The basic principle of the SCNN is to respond to the input mode between neurons in the competitive layer of the network. Only one neuron finally becomes the “winner” of the competition. The winning neuron expresses identification to the input mode. The connections of the self-organizing network between neurons of the competitive layer simulate the weights of the mutual inhibition phenomena of the neurons in the biological neural network layer. The inhibitory weights meet certain distribution relationships. The inhibition becomes stronger because of the closer distance but weakens when the distance is farther.

Assuming that the input layer consists of N neurons and the competitive layer consists of M neurons. The weight value of the network connection is W_ij, where i = 1, 2, …, M and the sum of W_ij equals one. The neurons mutually compete in the competitive layer. Thus, only one or several neurons adapt to the current input sample. The winning neurons represent the classification model of the current input sample. The input sample of the competitive network is a binary vector, and each element value represents zero or one. The state of neuron j in the competitive layer can be calculated as follows:

1. where, X_i is the first element of the input sample vector and neuron k with a maximum weight value in the comαpetitive layer. Thus, the input is as follows:

All “i” after competition is corrected according to the formula:

where, α refers to the learning parameter, 0 < α < 1 generally ranges from 0.01 to 0.03, and m refers to the number of neurons in the input layer with an input equal to one, m = X_i. In the weight adjustment formula, X_i/m indicates that the weight increases with X_i = 1, whereas the weight decreases with X_i = 0. Thus, the weight value of I increases when X_i is active but decreases when X_i is inactive. Given that the sum of the weight value equals one when the weight value of i increases or decreases, other weight values may decrease or increase in response. The formula also ensures that the adjustment of weights meets the following criterion: the sum of the adjustment values equals zero.

RESULTS AND DISCUSSION

NIR Spectra

The NIR spectra of the 180 samples comprising four kinds of pork (HM, CM, FM, and CFM) are shown in Fig. 2. The shapes of the NIR spectra are similar, and the four categories of pork samples cannot be directly distinguished.

FIGURE 2 Near-infrared spectra of the various pork samples.

PCA

The as-obtained spectra are influenced by high-frequency random noise, baseline drift, a non-uniform sample, and light scattering. Thus, the original NIR spectra cannot be used for any direct calculation. To eliminate the influence of the raw spectral curve, the as-obtained data were pretreated by four kinds of spectral preprocessing methods, namely, MSC, SNV, first-order differential, and second-order differential. The optimum spectral preprocessing method was determined by comparison.

Each spectrum of the pork sample contained 870 data points that were too many to calculate. The spectral information of some regions of the sample was weak and lacked correlation between sample components and properties. Thus, the spectrum data were analyzed with MATLAB after pretreatment. The cumulative contribution rate of the first four principal components is shown in Table 1. Considering that the accumulated credibility of the first four principal components was >95%, only the first four principal components represented the primary information of the original NIR spectra. Thus, the data arrays were reduced from 180 × 870 to 180 × 4.

TABLE 1 Accumulative contribution rate of the first four principal components

Principal component	PC1	PC2	PC3	PC4
Contribution rate (%)	43.42	36.74	13.11	1.89
Cumulative contribution rate (%)	43.42	80.16	93.27	95.16

PCA revealed that the first and second main components can be used as horizontal and vertical coordinates, respectively, to construct principal component score plots. These plots provided some references for the identification of sample and the choice of pretreatment method. Figures 3a–3d indicate the principal component score plots pretreated by MSC, SNV, first-order differential, and second-order differential, respectively.

FIGURE 3 A: Multiplicative scatter correction principal component score plot; B: Standard normal variables principal component score plot; C: First differential derivative principal component score plot; D: Second differential principal component score plot.

The clustering effect of the principal component score plot obtained from the spectrum data (pretreated by MSC, SNV, and first-order differential, respectively) was not ideal. The sample distribution was discrete and had an obvious overlap of the four categories of pork samples. The boundaries of sample species in the principal component score plot pretreated by the second-order derivative spectrum data pretreatment were relatively clearer but still showed a partial overlap. This overlap cannot be directly used to distinguish pork types from the principal component. Thus, based on the spectra pretreated by the second-order derivative, the analysis model was established by dimension reduction, i.e., through PCA combined with SCNN.

Modeling of SCNN

Calibration Models results are shown in Table 2. We noted 15 recognized errors from the observation of the 120 samples. The overall correct recognition rate was 87.5%, and the correct identification rate of CFM reached 90.0%. The network needed to be tested to determine the optimum learning number of the classification performance. The 120 samples were classified by different learning numbers, and the results are shown in Table 3. Table 3 shows that the highest rate of the overall recognition accuracy was 91.7% when the learning number of the SCNN was 250. The recognition accuracy rate also reached the highest at 96.7%, which indicated the optimum modeling effect. The remaining 60 samples (four categories, 15 samples each) were used as the prediction set. We tested the model in which the optimum learning number was 250, and the results are shown in Table 4. Only four recognition errors appeared in the 60 samples. The overall correct recognition rate was 93.3%, in which the repeatedly frozen identification rate was 100%.

TABLE 2 Calibration models results

True value	Predictive value				Total sample numbers	Correct rate (%)
	HM	CM	FM	CFM
HM	27	1	2	0	30	90%
CM	2	13	0	1	16	81.3%
FM	3	1	16	0	20	80%
CFM	0	1	1	14	16	87.5%

TABLE 3 Effect of learning numbers on the model

Learning numbers	HM	CM	FM	CFM	Correct recognition numbers	Correct recognition rate (%)
	Correct recognition numbers	Correct recognition numbers	Correct recognition numbers	Correct recognition numbers
100	27	26	25	27	105	87.50%
200	29	27	25	26	107	89.20%
250	28	27	26	29	110	91.70%
300	26	26	28	28	108	90.00%
350	25	29	25	26	105	87.50%
400	25	28	24	27	104	86.70%

TABLE 4 Testing results of the prediction set

True value	Predictive value				Total sample numbers		Correct rate (%)
	HM	FM	CM	CFM
HM	14	1	0	0	15	93.3%
FM	0	14	0	1	15	93.3%
CM	0	1	14	0	15	93.3%
CFM	0	0	0	15	15	100.0%

CONCLUSIONS

A total of 180 pork samples including four categories, namely, HM, CM, FM, and CFM were identified by NIR spectroscopy combined with SCNN. SCNN modeling analysis was adopted after the pretreatment of data by MSC, SNV, first-order differential, and second-order differential. The correct recognition rate of CPM reached 100%. Thus, the combined use of NIR spectroscopy and SCNN to identify CPM was effective. Given that NIR was fast, accurate, and inexpensive, it can conveniently and non-destructively monitor the quality of pork.

ACKNOWLEDGMENTS

The authors gratefully acknowledge many of their colleagues for their stimulating discussions in this field.

FUNDING

This work was supported by the Specialized Research Fund for the Doctoral Program of Higher Education of China (20124401120005), the Key Technologies R&D Program of Guangdong Province, China (2012A032300016), the Natural Science Foundation of Guangdong Province, China (S2011040001850), the Guangdong College of Outstanding Youth Innovation Talent Training Project in China (LYM11026), and the Fundamental Research Funds for the Central Universities, China (21612436 and 21612353).

Additional information

Funding

This work was supported by the Specialized Research Fund for the Doctoral Program of Higher Education of China (20124401120005), the Key Technologies R&D Program of Guangdong Province, China (2012A032300016), the Natural Science Foundation of Guangdong Province, China (S2011040001850), the Guangdong College of Outstanding Youth Innovation Talent Training Project in China (LYM11026), and the Fundamental Research Funds for the Central Universities, China (21612436 and 21612353).