Introduction

The pig industry plays a vital role in global meat production, which provides 33% of total meat consumption1 and currently has become the second-largest consumed meat worldwide. With the improvement of people’s living standards and consumption upgrades, the demand for high-quality pork has become an inevitable trend. Pork quality, a comprehensive concept, comprises a variety of attributes related to fresh pork and related processed products, while the quality of fresh pork traditionally can be evaluated by a set of properties, such as color, marbling, pH, drip loss, water-holding, tenderness, flavor, and odor. Among these indicators, meat color is the paramount attribute that significantly influences consumer perception and purchase decisions2, and discoloration can cause considerable economic losses3. Therefore, meat color has attracted wide attention from farms, slaughterhouses, retailers, and consumers, and the objective and rapid assessment of meat color has become increasingly important throughout the entire pig production process.

Meat color measurement methodologies are mainly divided into two categories, namely visual assessment and instrumental measurement, the details on meat color measurement methods were completely reviewed in the American Meat Science Association (AMSA) Meat Color Measurement Guidelines (http://www.meatscience.org, 2012). The traditional colorimeter measurement is widely used for meat color evaluation. It objectively characterizes meat color by collecting the feature parameters of L* (lightness), a* (redness), and b* (yellowness), hue angle (H0, discoloration), and saturation index or chroma (C*, red color intensity). However, the disadvantages of these colorimeters measurement are that the measured surface of meat must be uniform and the measurement area is rather small (2–5 cm2)4, and the grades of meat color cannot be obtained directly by the colorimeters and it must be judged subjectively according to the absolute values of feature parameters. Notably, there was a certain difference in the feature parameters of color measured by various colorimeters5, suggesting that the final output of meat color grades may be inconsistent in different instrumental measurement systems.

Visual evaluations of meat color are the “fundamental standard” of color measurements because they closely relate to consumer perception evaluations and are the benchmark for instrumental measurement comparisons. Visual evaluation of meat color is influenced by a variety of conditions, such as type and angle of illumination, environmental differences, and human factors, but objective visual appraisals of meat color can be obtained by the well-trained panelists under a standardized condition. Moreover, the standard reference materials, such as the standard board for color and marbling of pork constructed by the National Pork Producers Council (NPPC) in 1999, are usually used to enhance the reliability of visual meat color by subjective evaluation. In spite of this, visual sensory evaluation still has some unavoidable limitations, such as the low efficiency of evaluation by trained panelists or consumers, prone to visual fatigue, and the accuracy is often influenced by the sensitivity of the individual’s eyes6.

In recent decades, some novel technologies for meat quality evaluation have emerged, characterized by different advantages, including being rapid, consistent, objective and non-invasive7,8,9. Among these, the computer vision system (CVS) has the advantage of providing more accurate measurements, closer to the visual color, and these advantages have been validated in multiple species10,11,12,13, thus CVS is a promising method to obtain the precise meat color values. It is particularly worth noting that the rapid and accurate models for the evaluation of meat color scores are still lacking. Traditionally, the prediction models for the evaluation of meat color scores are developed based on large-scale samples using the feature parameters describing the meat color measured by instruments or other image features, and the performance of a model can be affected by the source of an animal population. For example, sun et al. (2018) conducted a model development for meat color scores using 1400 pigs from seven different plants, and the classification accuracy of developed models based on image color and texture features in different pig populations varied from 75.0%-92.5% with an average accuracy of 78.9%14. Several limitations have been found in the traditional method, such as a narrow scoring range of collected pork usually with 3.0 to 5.0 score on a 6-point scale, a large but unevenly distributed sample size, and the slope and intercept calibration depend on the source of the pork and device. To address these limitations, a novel method for rapidly constructing accurate pork color scoring models based on the NPPC standard color boards was proposed. The objectives of this study were to provide a quick method to construct pork color scoring models, which characterized by full-range coverage, uniform sample distribution, simple parameter calibration, and the slope determination of models is independent of pork sample source, and to verify the accuracy and robustness of the models. The main contributions of this study are summarized as follows:

  1. (1)

    Develop full-range coverage pork color scoring models based on the standard color board, and the slope of unbiased predictive model is independent of the sample source.

  2. (2)

    Fast intercept calibration of the model only depends a small of samples, thus making it easy to popularize in practice.

  3. (3)

    The pork color scoring models based on the standard color board achieve satisfactory classification accuracy, and the classification ability is similar to the models developed by the traditional method.

Results

Overview of the data

In this study, 525 pigs were slaughtered in three plants, involving in seven pig herds. The artificial color scores of pork derived from the experimental pigs are ranged from 3.0 to 5.0, but slight differences are observed in different pig herds (Table 1). The most pork color score is clustered at 4.0, followed by 3.5 (Fig. 1b). It shows that 9.90% of 3.0 score, 31.05% of 3.5 score, 43.05% of 4.0 score, 16.00% of 4.5 score, and 1.90% of 5.0 score. Among these pig herds, because the number of pigs of the LM and Me pigs was small, thus they were not used for subsequent classification accuracy analysis.

Fig. 1: Pork collection and artificial color scoring.
figure 1

a Flowchart of pork collection, (b) Flowchart of artificial color scoring and histogram of color score distribution.

Full size image
Table 1 Basic information on experimental pigs and pork scores
Full size table

Performance evaluation of pork color scoring models calibrated by the mixed pig herd

The performances of the pork color scoring models constructed using different image feature parameters are shown in Table 2. It shows that a high level of performances were achieved only using the L* feature parameter based the standard board images, and performance of the CS_1_L* model (R2 = 0.97) is comparable to that of CS_1_L*a* (R2 = 0.97)and CS_1_L*a*b* (R2 = 0.96), indicating that the performance of the models based the standard board images is less affected by the number of the image features parameters and L* feature parameter is critical for the performance of the models. Moreover, the results show that the performances of CS_2 models are consistent with those of CS_1 models, while the performances of CS_3modelsarefluctuated seriously with a R2 ranged from 0.50 to 0.81 as the increasing of image features parameters, and the higher performance was observed in the CS_3_L*a*b* model (R2 = 0.81), indicating the performances of the models constructed using the image features parameters from the real pork samples based the traditional method is significantly affected by the number of the feature parameter.

Table 2 Performance comparison of the pork color scoring models based on different image feature parameter
Full size table

Due to the differences of image features between the pork standard color board and the real pork sample images, so the models were further calibrated using the image feature parameters of the real pork samples derived from the mixed pig herd. Notably, the intercepts of the CS_1 model constructed based on the standard board images were changed greatly after calibration, while no obvious changes were observed in the other models constructed based on the real pork sample images(Table 2), indicating that it is necessary to calibrate the intercept using the real pork sample images when applying the models based on the standard board image. Although the intercepts were not changed greatly after intercept calibration for the models constructed using the feature parameters of the real pork sample images, which does not mean that these models have the high-performance characteristic. On the contrary, the stability of these models constructed using the images features of the real pork samples may be unsatisfied. For example, Sun et al. (2018) found that the overall prediction accuracy was 78.9% of model constructed using the image color and texture features of the pork from seven plants for pork color scores, while the prediction accuracies of the models were ranged from 75.0% to 92.5% in individual plant14, indicating the performances of models developed were affected by many external factors. In this study, the novel method proposed for constructing the CS_1 model has several characteristics, including full-range coverage, modeling with a small number of samples and balanced sample size, and the performance of the models is less affected by sample characteristics and ease to be calibrated with a small number of real pork samples.

Classification accuracy evaluation of the models calibrated by mixed pig herd at different pork color scores

In this study, three methods were used to construct the pork color scoring models, three models for each method (Table 2). To test the predictive ability of these models, the classification accuracy of all the calibrated models were firstly evaluated at different pork color scores with a scoring scale of ≤±0.25 and ≤±0.50.The results shows that the overall classification accuracy with a scoring scale ≤ ±0.50 is about 30% higher than that with a scoring scale ≤±0.25 for each model (Tables 3, 4), that is because artificial pork colors coring was performed based on a scale of ±0.50. Notably, the high classification accuracy with a scoring scale of ≤± 0.50 was observed for most of the models, but the obvious differences of classification accuracy were observed at different pork color scores in a specific model (Table 4). Moreover, the highest classification accuracy was mainly observed at a score of 5.0, but the results need to be further validated by increasing the number of testing samples. While the stably high classification accuracy was achieved at a score of 4.0 for all the models. The main reason is that the samples are dominated in a score of 4.0. In addition, the classification accuracies of most models were improved as the increases of the image feature parameters, and the classification accuracy of the models based on the image feature parameters of the pork standard color board images have the similar predictive ability with the models based the image feature parameters of the real pork sample images (Table 4).

Table 3 Classification accuracy of the models calibrated by mixed pig herd within a scoring scale ≤ ±0.25 at different pork color scores
Full size table
Table 4 Classification accuracy of the models calibrated by mixed pig herd within a scoring scale ≤ ±0.50 at different pork color scores
Full size table

Classification accuracy evaluation of the models calibrated by mixed pig herd in each pig herd

To explore the effect of different pig herds on the classification accuracy, the classification accuracy evaluation of the calibrated models was conducted in each pig herd. As shown Table 5 and Table 6 with a scoring scale of ≤±0.25 and ≤±0.50, respectively. Similarly, the results shows that the classification accuracies with a scoring scale of ≤±0.50areobviously higher than that with a scoring scale of ≤±0.25 for each model. However, the classification accuracies are fluctuated dramatically among the pig herds (Tables 5, 6). In the models constructed using the L* feature parameter, the classification accuracies are ranged from 65.96% to 96.94%; In the models constructed using the L* and a* feature parameters, the classification accuracies are ranged from 85.11% to 100.00%; In the models constructed using the L*, a* and b* feature parameters, the classification accuracies are ranged from 80.85% to 100.00%. Moreover, the overall highest classification accuracy was observed in the CS_3 models, but the CS_1 and CS_2 models have the comparative predictive ability with CS_3 models in a condition of the same method, and the CS_1 and CS_2 models show more similar predictive ability. Generally, these results indicate that the classification accuracies for all the models varied across different pig herds.

Table 5 Classification accuracy of the models calibrated by mixed pig herd within a scoring scale ≤ ±0.25 in different pig herds
Full size table
Table 6 Classification accuracy of the models calibrated by mixed pig herd within a scoring scale ≤ ±0.50 in different pig herds
Full size table

Classification accuracy evaluation of the models calibrated by individual pig herd in each pig herd

Due to the classification accuracy of the models is affected by the pig herds, so the calibration of the models using the image features of the real pork samples derived from the individual pig herd was subsequently performed and the classification accuracies of these calibrated models in each pig herd were evaluated. As shown in Table 7, the calibrated intercepts were different in each pig herd for the models. Notably, when compared with the calibrated intercepts in Table 2, the greater the change of calibrated intercepts (Table 7), the greater the corresponding classification accuracy will be improved for a specific model with a scoring scale of ≤±0.25 or ≤±0.5 (Table 5 vs. Table 8, Tables 6, 9). As shown in Table 9, the overall classification accuracies of all the models are improved greatly, and the classification accuracies of all the models constructed using the L*, a*, and b* feature parameters reach more than 96.00% and the models constructed by the three methods shows the similar predictive ability. These results indicate that the classification accuracies of all the models can be improved by the calibration of intercept, and a much larger improvement can be achieved using the novel model construction method based on the pork color standard boar proposed in this study compared to the traditional method, and the similar predictive ability of the models constructed by the novel method can be obtained after the intercept calibration compared to the models constructed by the traditional method.

Table 7 Intercept calibration of pork color scoring models using the feature parameters of real pork samples from individual pig herd
Full size table
Table 8 Classification accuracy of models calibrated by individual pig herd within a scoring scale ≤ ±0.25 in each pig herd
Full size table
Table 9 Classification accuracy of models calibrated by individual pig herd within a scoring scale ≤ ±0.50 in each pig herd
Full size table

Discussion

In this study, a novel method for the rapid construction of a precise color scoring models based on the pork color standard board images was proposed, and the classification accuracies of the models developed by this novel method are comparable to the models constructed by the traditional method across all pork color scores. For example, under a scoring scale of ≤±0.50, all the overall classification accuracies of all the models calibrated by mixed pig herd based on L*, a* and b* feature parameters reaches more than 94.00%. In the pork color scoring models, the evaluable range of models is influenced by the pork color scoring range of the samples used to develop the models. However, most of the pork collected were distributed between 3.5 and 4.0 color score in this study, with an overall range of 3.0 to 5.0 color score. In contrast, a completed distribution of 1.0 to 6.0 color score can be provided by the pork color standard boards. Therefore, the color scoring models constructed based on the pork color standard boards can be adapted to the full range of pork color evaluation. Moreover, the results showed that the models achieved higher classification accuracy for pork color score in a condition of large number of samples than that in a condition of small number of samples.

CS_1 model was developed only using the feature parameters of six pork color standard board images and the intercepts were calibrated only using five real pork samples, while525pork was needed to determine the slope and intercept of the CS_2 and CS_3 models. Therefore, the novel method proposed for the construction of CS_1 model exhibits the advantages of simple model construction and fast parameter calibration.

The pig herd is critical factor affecting the classification accuracy of the models. The results show that the overall classification accuracies are greatly improved in a scoring scale of ≤±0.25 and ≤±0.50 by the intercept calibration using the feature parameters of the real pork sample images derived from individual pig herd, and the differences of the classification accuracy among pig herds are reduced obviously. Moreover, the results show that a relatively small improvement of classification accuracies were observed a scoring scale of ≤±0.50 when compared to the scoring scale of ≤±0.25 after intercept calibration in different herds, that is because the relatively high classification accuracy of the models already were existed an uncalibrated model. Together, the intercept calibration is an effective method to enhance the classification accuracy and it should be performed for individual pig breed.

In this study, most of pork collected were concentrated in the 3.0–4.5 color score, the pork with the color scores at 2.0 or below and 5.0 or above are not enough. In the future, it will be necessary to increase the amount of pork with the color scores of 2.0 or below and 5.0 or above to improve the data set, and further evaluate the applicability of color scoring models for specific scoring levels. Meanwhile, there were significant differences in sample size among the different pig herds in this experiment, and a larger sample size could potentially confound the effects of genetic factors on classification accuracy. In the future, we will conduct in-depth research on the individual effects of pig herd and sample size on the model, as well as their interactive effects.

Methods

Animals

In this study, a total of 525 pigs (the information of them listed in Table 1) involving, in seven pig herds were slaughtered in three plants. Specifically, 294 castrated Large White pigs (LW) boars and 89 castrated Landrace × Large White pigs (LL) boars were from plant1, 32 castrated Pietrain × Meishan pigs (PM) boars, 10 castrated Large White × Meishan pigs (LM) boars, 3 castrated Meishan pigs (Me), and 47 Suhuai pigs (Su) including 27 castrated boars and 20 gilts were from plant 2, and 50 Duroc × (Landrace × Large White pigs) (DLL) including 24 castrated boars, 22 gilts, four pigs undocumented sex were from plant3. Before slaughter, the pigs were fasted and allowed to rest and drink freely for more than 12 h, but the feeding water was stopped 3 h preslaughter. The pigs were immobilized by electrical stunning and slaughtered humanely at a fixed standardized commercial abattoir according to the Chinese national standard (GB/T 17236-2019). All protocols involving animals were approved by the Animal Protection and Utilization Committee of Nanjing Agricultural University (Certificate No.: Code: SYXK (Su) 2022-0031).

Sample collection for visual evaluation

Artificial scoring of real pork color was conducted according to the agricultural industry standard of the People’s Republic of China, Technical Procedures for Pork Quality Determination (NY/T821-2019). Briefly, at a room temperature of 4 °C, ~2 cm thick cross-section of longissimus dorsi muscles at the third and fourth last thoracic ribs with an area of around 40 to 70 cm² were firstly dissected from the left half of the carcass at 45 minutes after the pigs slaughtered. Then, the pork was placed on a panel with a white background, and the real pork color were assessed and recorded by a fixed professional assessor according to the NPPC standard color board with a range of 1.0 to 6.0 grades, and 0.5 scale was allowed in the process of color grading, which was expressed as CS. The flowchart of artificial color scoring is shown in Fig. 1.

Images acquisition of the samples

The images of real pork samples were acquired immediately using the scanner after the artificial pork color scoring was completed. Meanwhile, the images of standard color boards were collected in the same way as the images of the real pork sample. The image acquisition process was shown in Fig. 2(a1), (b1). In this study, The EPSON V370 scanner with an LED light source with a color temperature of 6500 K was used, and the resolution of the image acquisition was set as 400 dpi. To eliminate the influence of reflected light during scanning, the inside of the cover board was covered with a black photographic cloth. Moreover, the scanner was subjected to color calibration using the IT8 color card before images acquisition15.

Fig. 2: Acquisition and processing of image.
figure 2

a1 Flowchart of color standard board image acquisition, (b1) Flowchart of pork image acquisition, (a2) Color standard board image processing procedure, (b2) Real pork sample image processing procedure.

Full size image

Extraction of color coordinates

The L*, a*, and b* values are feature parameters commonly used to evaluate color and are independent feature parameters in Lab color space. L*, a*, and b* indicate the lightness, red-green chromaticity, and blue-yellow chromaticity of the sample, respectively. To eliminates the influence of background and white adipose tissue at the edges of pork on the calculation of L*, a*, and b* values, the background of the images and white adipose tissue at the edges were removed using the MATLAB 2020 software. The process of image was shown in Figs. 2(a2), (b2). Subsequently, the L*, a*, and b* values of the processed images derived from the standard color boards and the real pork samples were calculated. In this study, the images of six standard color boards and 525 real pork samples were processed to obtain the L*, a*, and b* values, which provides fundamental data for the construction of pork color scoring model.

Modeling

Three methods were used to develop the pork color scoring models, one was based on images of flat-surfaced NPPC standard color boards with a color score range from 1.0 to 6.0, the others were based on the real pork sample images. The diagram of the above methods is shown in Fig. 3. The process was divided into three main steps. First, model construction involves establishing a ridge regression equation between the feature parameters and the pork color scores. The modeling samples should cover the full range of pork color scores from 1.0 to 6.0 as comprehensively as possible. Second, model calibration refers to the rapid determination of the slope and intercept of the regression equation, with a clear understanding of their physical meanings. The samples for calibration should be as standardized and accessible as possible. In this study, the images of six pork color standard boards or a large number of real pork samples were used to determine the slope of the models, while five pork samples with a score of 3.5 were randomly selected for intercept calibration. The reason for using pork samples with a score of 3.5 is that such samples are relatively easy to obtain. Actually, using a small number of samples allows for quick calibration of the model parameters, facilitating its application in practice16. Third, model evaluation involves comparing the classification accuracy of pork color scoring models across different pig herds and color scores. This will help us analyze the factors affecting model performance and assess the applicability of the novel model construction method.

Fig. 3
figure 3

A schematic diagram of the methodology for rapid construction of precise pork color scoring model.

Full size image

The specific three methods used to construct the pork color scoring models were provided as follows:

  1. (1)

    The first method for constructing a pork color scoring model (CS_1) was based on the feature parameters of standard color board images. The ridge regression method in MATLAB software was used to construct the models using the feature parameters of the standard color board images as independent variable and the corresponding 1.0 ~ 6.0 scoring values of the standard color boards as the dependent variable. When L* value was used in the model construction, the fitted equation was denoted as CS_1_L*; when L* and a* values were used in the model construction, the fitted equation was denoted as CS_1_L*a*; when L*, a*, and b* values were used in the model construction, the fitted equation was denoted as CS_1_L*a*b*.

  2. (2)

    The second method for constructing a pork color grading model (CS_2) was based on the feature parameters mean value of the real pork sample images. Similarly, the ridge regression method in MATLAB software was used to construct the models using the feature parameters mean value of the real pork sample images as independent variable and the mean grading value of the real pork sample as the dependent variable. Specifically, the L*, a*, and b* feature parameters of all pork images were firstly classified according to the artificial grades ranged from 1.0 to 6.0. Then, the means of L*, a*, and b* of pork images at the same scores, as well as the corresponding artificial score, were used to construct the models. When the L* mean value was used in the model construction, the fitted equation was denoted as CS_2_L*; when the L* and a* mean values were used in the model construction, the fitted equation was denoted as CS_2_ L*a*; and when the L*, a*, and b*mean values were used in the model construction, the fitted equation were denoted as CS_2_L*a*b*.

  3. (3)

    The third method for constructing a pork color grading model (CS_3) was based on the feature parameters of real pork sample images. Similarly, the ridge regression method in MATLAB software was used to construct the models using the feature parameter of the real pork sample images as independent variable and the scoring values of the real pork sample as the dependent variable. Specifically, the L*, a*, and b* feature parameters of each pork image, along with the corresponding artificial pork color scores, were used to construct the models. The model including the L* feature parameter was denoted as CS_3_L*; The model including the of L* and a* feature parameters were denoted as CS_3_L*a*; while the model including the L*, a*, and b* feature parameters were denoted as CS_3_L*a*b*.

In this study, the coefficient of determination (R²) and root mean square error (RMSE) of the fitted equations was used to evaluate the performance of models constructed using the three aforementioned methods.

The specific model calibration and model evaluation were conducted as follows:

After the fitted equations were constructed, five pork images with an artificial score of 3.5 were randomly selected to calibrate the intercepts of the fitted equations. Specifically, five pork images with an artificial score of 3.5 were processed according to the method described in the Section of extraction of color coordinates and their L*, a*, and b* feature parameter values were obtained. Then, these values were input into the constructed fitted equations to calculate the five intercepts for each equation. The average value of five intercepts was used as the calibrated intercept, and the intercepts in the aforementioned fitted equations were replaced with the calibrated intercept to obtain the calibrated regression equations. Furthermore, the pork color scores of all the remaining pork images were predicted using the calibrated regression equations, and the classification accuracy was evaluated by comparing with the artificial pork color scores. Classification accuracy was assessed based on the thresholds of ≤0.25 and ≤0.50, and the effects of different scores, herds on the classification accuracy of pork color scoring models were further evaluated. The accuracy calculation for the classification models was given by the following Eq. (1):

$${\rm{Accuracy}}=\frac{{\rm{Num}}({\rm{ACS}}-{\rm{MCS}}\le {\rm{Th}})}{{\rm{Num}}({\rm{Total}})}\times 100 \%$$
(1)

Where ACS represents the artificial pork color score, MCS represents the model’s pork color score, and Th represents the threshold values of 0.25 and 0.50, respectively.