Prediction of total volatile basic nitrogen (TVB-N) and 2-thiobarbituric acid (TBA) of smoked chicken thighs using computer vision during storage at 4 °C

Highlights

• Color data of smoked chicken thighs obtained by computer vision during storage.

• Multiple linear regression models for predicting TVB-N and TBA fitted well.

• Distribution maps were generated to predict the value of TVB-N and TBA directly.

• Distribution map can be used as a fast and intuitive freshness prediction method.

Abstract

As the traditional indicators of freshness measurement of meat products, TVB-N and TBA have the disadvantage of time-consuming, labor-intensive and destructive to the sample. The objective of this study was to investigate the possibility of computer vision techniques to visualize the variation of TVB-N and TBA during the storage of smoked chicken thighs. In this study, freshness indicators (TVB-N and TBA) and images of smoked chicken thighs were obtained simultaneously every 3 days during storage at 4 °C. Then, the RGB color space was converted to HSI and L*a*b* color spaces by color conversion algorithm, and the color parameters (RGB, HSI and L*a*b*) were correlated with TVB-N and TBA, respectively, for establishing multiple regression models. Finally, visualization maps of the spoilage were established by applying the multiple regression model to each pixel in the image. The results showed that the multiple linear regression models of TBA and TVB-N based on the color parameters L*, a*, I, S and R were well correlated (R ² = 0.993 for TBA and R ² = 0.970 for TVB-N). Distribution maps of TBA and TVB-N changed color gradually from blue to red during storage, respectively. In conclusion, this study demonstrated that distribution maps can be employed as a rapid, objective, and non-destructive method to predict the TBA and TVB-N values of smoked chicken thighs during storage.

Introduction

Smoked chicken thighs, as one of the traditional foods of smoking meat, have the characteristics of unique smoky aroma (Zhang et al., 2021a) and appealing smoky color (Wang et al., 2021a, Wang et al., 2021b, Zhang et al., 2021b), which are popular among numerous customers. However, protein dehydration, denaturation, degradation, and lipid oxidation may occur during storage or industrial processing due to the rich nutritional properties of smoked chicken thighs, which will affect its freshness (Lan et al., 2016). Therefore, freshness, as one of the most important factors in evaluating the quality and safety of meat, is highly correlated with the sales and consumption of smoked chicken thighs.

In general, there are two main methods to assess the freshness of food products: subjective measurement and objective measurement. Sensory analysis, as a subjective measurement method, is conducted by the assessors and involves the use of eyesight, touch, and olfaction (Karpinska-Tymoszczyk, 2014). Objective measurement methods include physical methods such as texture (Dang et al., 2021), color (Lan et al., 2016) and electronic nose (Bekhit et al., 2021); chemical methods such TBA content (Karpinska-Tymoszczyk, 2014) and TVB-N content (Liu et al., 2020, Lan et al., 2016); biochemical methods such as trimethylamine (TMA), biogenic amine formation (Bekhit et al., 2021) and microbiological methods such as flat colony counting method (Wang et al., 2020, Shange et al., 2019). These objective analytical methods are more accurate than sensory analysis and play a critical part in the current methods of freshness evaluation. However, a considerable amount of time and energy is required to carry out these traditional determination methods, and it can also be destructive to the sample (Shi et al., 2018). Meanwhile, the analytical reagents used in the experiments are potentially harmful to the laboratory staff and the environment (Dallinger et al., 2020). Therefore, developing an objective, rapid and non-destructive method to evaluate the freshness of smoked chicken thighs would be significant to the smoked chicken industry due to its potential economic effect.

Computer vision has been applied as a fast and non-destructive method to evaluate food quality by analyzing color changes (Saldaña et al., 2014). It is also a critical technique for evaluating food freshness (Shi et al., 2018). It has the advantages of efficiency, objectivity, consistency and reliability. Recently, there are more and more researches using computer vision evaluates the quality of chicken (Taheri-Garavand et al., 2019), fish (Dowlati et al., 2013) and pork (Saldaña et al., 2014) by color features. However, meat spoilage is a complicated and dynamic process due to the tissue enzymes and microorganisms. It is common for meat degradation that accompanies simultaneous changes in internal properties (chemical composition) and external properties (color, texture and odor) (Shi et al., 2018, Lan et al., 2016). According to You et al. (2020), lipid oxidation and protein decomposition during storage will accelerate myoglobin oxidation and eventually led to the color change of meat. By establishing the multiple linear regression model. Zhang et al. (2019) determined that color was highly correlated with the freshness index of chilled lamb. The freshness of red mullet was evaluated using the image processing tool associated with chemical and sensory analysis by Tappi et al. (2017).

However, traditional methods for predicting the content of components can only obtain the average value of the measured parameters and cannot provide their space distribution. Meanwhile, due to the fact that the location distribution of each pixel in the sample image was spatial, it provided the possibility to map the spatial distribution of different physicochemical components. For example, Cheng et al. (2016) performed pseudo-color visualization of the established partial least squares regression (PLSR) model with the biogenic amine index corresponding to each pixel in the hyperspectral image to achieve the mapping of pork biogenic amine index on the spatial image. Shi et al. (2018) proposed that the freshness of tilapia could be represented visually as a distribution map based on images acquired by computer vision. Torres et al. (2021) established a distribution map of almond fatty acids through the established partial least squares (PLS) regression model and hyperspectral images. Meanwhile, Wang et al., 2021a, Wang et al., 2021b achieved not only the prediction of tea polyphenol content in each pixel of the tea spectral image by transferring the filtered optimal prediction model to each pixel of the hyperspectral image, but also visualized the spatial distribution of tea polyphenols in tea samples by the established distribution map. However, there were no studies reported to obtain the spatial distribution of TBA and TVB-N during the storage of smoked chicken thighs directly by distribution maps.

The objective of this study was to investigate the feasibility of distribution maps as a rapid and non-destructive method for predicting the freshness of smoked chicken thighs based on computer vision technology. It provides a reference for future research on the use of computer vision to predict other freshness parameters and may also provide a way for the meat industry to achieve real-time monitoring of changes in the quality of smoked chicken products.

In order to benchmark the actual production of the company and reduce the effect of freezing on the quality of raw meat (Qi et al., 2021), 390 (160 ± 10 g) chicken thighs were provided by Liaoning Goubanzi Food Co., Ltd (Liaoning, China) with a freezing time of less than two weeks. The unpeeled chicken thighs were about 5.5 cm in diameter. All the chicken thighs were transferred to the Lab at Bohai University in a cooling box at 2–4 °C. Sucrose was purchased from Nanjing Ganjuyuan Sugar Industry Co., Ltd. (Nanjing, China). All spices were purchased from a supermarket located in Jinzhou, Liaoning Province, China. Neutral filter paper, packaging bag (18 × 25 × 0.016 cm), anhydrous alcohol, methyl red and all other chemical reagents of analytical grade were supplied by Jinzhou National Pharmaceutical Co., Ltd. (Jinzhou, China).

The preparation of sugar-smoked chicken thighs was referred to the method described by Wang et al., 2021a, Wang et al., 2021b. The technology of sugar smoking was as follows: Chicken thighs were thawed under running water at room temperature (25 ± 5 °C) and were blanched. Boiled chicken thighs in brine for 10 min, and then simmered for 90 min. The recipe for the spices of the brine was prepared according to the method of Wang et al., 2021a, Wang et al., 2021b. Then, remove the chicken thighs from the brine and chill it to room temperature (25 ± 5 °C).

Sucrose (350 g) was added and began to produce smoke when the temperature of the sugar heating plate in the sugar fumigation furnace (YXDT1/1, Hebei Xiaojin Machinery Manufacturing Co., Ltd., Hebei, China) with a length, width and height of 152 × 130 × 260 cm, respectively, increased to 330 °C. When the furnace temperature reached 100 °C, the chicken thighs were hung in the sugar fumigation furnace and sugar smoking for 8 mins. The sucrose amount used in each batch was equal. When the temperature cooled to 15 ± 5 °C, the oil stains were removed from the surface of the chicken thighs. Afterwards, each sugar-smoked chicken thigh was packed in vacuum (DZ-600, Jinda, China) tagged with the date and sequence number with storage in a refrigerator at 4 °C for 36 days. Finally, collect chicken thighs images and determine the TVB-N value and TBA value every three days for 36 days. A total of 13 groups of sample images and corresponding TVB-N value and TBA value were obtained.

The image acquisition system is shown in Fig. 1. The image collection system was composed of an imaging chamber (Pangniu Technology Co., Ltd., Shenzhen, China), a bracket (ES400300, LOTS, Dongguan, China), a digital camera (EOS 600D, Canon, Tokyo, Japan), a computer (Y7000P, Lenovo, Beijing, China), a ring LED shadowless light (HPR200, LOTS, Dongguan, China) for image processing. The imaging chamber was 50 cm in length, width and height. A ring shadowless LED lamp (power of 15.2 W and color temperature of 6500 Kelvin) with an external diameter of 21.5 cm and an internal diameter of 17 cm was fixed in the middle of the top of the imaging chamber. This ring shadowless LED lamp is a planar light source, which can maximally reduce the reflection problem in the process of taking images. The bracket was placed inside the imaging chamber. The sample support plate was fixed to the bracket and can be moved vertically to adjust the distance between the sample and the camera, which minimizes the reflection and color deviation caused by the unsuitable light distance. The camera lens was placed in a round hole of 6 cm, which was created in the center of the top of the imaging chamber.

The Python program was applied to obtain and analyze the color information of the sugar-smoked chicken thighs in three steps. First, images were acquired by a digital camera for the smoked chicken thighs. Secondly, the R, G and B information of the image was obtained by computer vision. Finally, the RGB values were transformed into HSI and L*a*b* values by image analysis algorithms.

As the purpose of this study was to achieve nondestructive detection of spoilage degree of smoked chicken thighs, thus images were collected from vacuum-packed smoked chicken thighs. The digital camera was positioned vertically above the sample plate with the lens at a distance of 40 cm from the sample surface. This distance not only obtains a satisfactory image, but also minimizes the reflection problem caused by vacuum packaging. A total of 30 smoked chicken thighs were selected randomly from the refrigerator every 3 days to obtain pictures. The sample was flipped 90° after acquiring one image, and four images were acquired in total for one sample. The external humidity of the sample was wiped before image acquisition. It was RGB color images that were collected in the experiment. The resolution of these images was 5184 × 3456 pixels and the format was PNG. To better display the effect of image preprocessing and obtain the local image from the whole smoked chicken thighs, the resolution of the image was reduced to 384 × 384 pixels. After the image was rotated, flipped and stretched randomly, the three new images were generated. It changed the position and orientation of the original image without changed the image properties. The number of sample images was increased and the accuracy of the test results was enhanced by preprocessing the images.

It was common that the color spaces RGB (red, green, blue), HSI (hue, saturation and intensity) and L*a*b* (lightness, redness and yellowness) were applied in food classification (Hashim et al., 2011). As a linear color space, RGB color spaces can be acquired rapidly through computer vision systems. Color differences on a uniform scale cannot be represented by RGB space only (Cheng et al., 2001). Thus, it does not easily distinguish the similarity of two colors using the RGB values. The HSI color space can help overcome the speed limitations related to color-based computer vision compared to the RGB color space (Gunasekaran, 1996). In contrast to the RGB and HSI color spaces, the L*a*b* color space contains a broader color gamut, which indicates that L*a*b* can exhibit a broader range of colors (Yam and Papadakis, 2004). The RGB color space was converted to HSI and L*a*b* color spaces to better describe the color information of smoked chicken thighs (Łuszczkiewicz-Piątek, 2014).

As described by León et al. (2006), the RGB color space was converted to L*a*b* color space in two steps.

Firstly, RGB values were converted to XYZ values.

$$\text{X}\text{=}\text{0.412453R}\text{+}\text{0.3578G}\text{+}\text{0.180423B}$$

$$\text{Y=0.212671R+0.71516G +0.072169B}$$

$$\text{Z=0.019334R +0.119193G+0.950227B}$$

Second step, convert XYZ values to L*a*b* values.

$${\text{L}}^{\text{*}}\text{=116f}\left(\text{Y}/{\text{Y}}_{\text{n}}\right)\text{-16}$$

$${\text{a}}^{\text{*}}\text{=500}\left[\text{f}\left(\text{X}/{\text{X}}_{\text{n}}\right)\text{-}\text{f}\left(\text{Y}/{\text{Y}}_{\text{n}}\right)\right]$$

$${\text{b}}^{\text{*}}\text{=200}\left[\text{f}\left(\text{Y}/{\text{Y}}_{\text{n}}\right)\text{-f}\left(\text{Z}/{\text{Z}}_{\text{n}}\right)\right]$$

$$\text{f}\left(\text{t}\right)\text{=}\left\{\begin{array}{c}\sqrt[\text{3}]{\text{t}}\text{i}\text{f}\text{t}>{\left(\frac{\text{6}}{\text{29}}\right)}^{\text{3}}\\ \frac{\text{1}}{\text{3}}{\left(\frac{\text{29}}{\text{6}}\right)}^{\text{2}}\text{t}\text{+}\frac{\text{4}}{\text{29}}\text{i}\text{f}\text{t}\le {\left(\frac{\text{6}}{\text{29}}\right)}^{\text{3}}\end{array}\right.$$

where X_n = 95.047, Y_n = 100.0, Z_n = 108.883, t refers to f(X/X_n), f(Y/Y_n), f (Z/Z_n).

The L* a* b* values of each storage period were compared with day 0 to determine the change in color difference (ΔE) of smoked chicken thighs in the whole storage period.

$$\mathrm{\Delta }\mathrm{E}=\sqrt{{\left({{\text{L}}_{\text{0}}}^{\text{*}}\text{-}{{\text{L}}_{\text{i}}}^{\text{*}}\right)}^{\text{2}}\text{+}{\left({{\text{a}}_{\text{0}}}^{\text{*}}\text{-}{{\text{a}}_{\text{i}}}^{\text{*}}\right)}^{\text{2}}\text{+}{\left({{\text{b}}_{\text{0}}}^{\text{*}}\text{-}{{\text{b}}_{\text{i}}}^{\text{*}}\right)}^{\text{2}}}$$

where ΔE represents the absolute color difference; L₀*, a₀*, b₀* represents the color parameters on day 0; L_i*, a_i*, bi* represents the color parameters on day i.

The RGB values were transformed into HSI values according to the method of Rotaru et al. (2008).

$$\text{I=}\frac{\text{R+G+B}}{\text{3}}$$

$$\text{S=1}-\frac{\text{3}}{\text{R+G+B}}\times \text{min}\left(\text{R, G, B}\right)$$

$$\text{H=}{\text{cos}}^{\text{-1}}\left(\frac{\text{0.5}\times \left[\left(\text{R-G}\right)\text{+}\left(\text{R-B}\right)\right]}{\sqrt{{\left(\text{R-G}\right)}^{\text{2}}\text{+}}\left(\text{R-B}\right)\left(\text{G-B}\right)}\right)$$

TVB-N was assessed by the semi-micro Kjeldahl method used by Hu et al. (2013). The analytical conditions were as follows: The method was as follows: distilled water (100 mL) was added into the minced sample (10 g) and then stirred with an electronic stirrer (30 min) (Allegra 64R, Thunder Magnetic Instrument Co., Ltd., Shanghai, China). Then, after centrifugation at 5500 r for 15 min (USA.JB-2A, Thunder Magnetic Instrument Co., Ltd., Shanghai, China), the mixture was filtered through ordinary filter paper, and the filtration was repeated twice to collect the permeate. Afterwards, 5 mL of the resulting filtrate and 5 mL of MgO suspension (10 g/L) were added into the tube, with 10 mL of boric acid (20 g/L) as the acceptor and then distilled using a Kjeldahl Apparatus (K9840, Haineng Scientific Instrument Co., Ltd, Shandong, China). The boric acid absorbent solution was titrated by 0.01 mol/L hydrochloric acid (HCl).

The TBA values were measured by the approach proposed by Lan et al. (2016). The minced sample (10 g) was mixed with 20% trichloroacetic acid (TCA) (w/v) solution (25 mL) and homogenized at 10,000 rpm for 30 s (FW-200, Zhongxingweiye Instruments Co., Ltd., Beijing, China). The supernatant was filtered twice after centrifugation at 5500 rpm for 15 min at 4 °C (Allegra 64R, Thunder Magnetic Instrument Co., Ltd., Shanghai, China). The filtrate (2 mL) and 0.02 mol/L TBA (2 mL) were mixed in a test tube and heated in a water bath at 80 °C for 20 min. Then, the solution was cooled 10 min in an ice bath. The absorbance of the sample was determined at 532 nm by using water as a blank (UV2550, Spectrum General Technology Co., Ltd., Beijing, China). A control sample was prepared and treated with the same procedure.

Multiple linear regression modeling was a statistical technique that was used to model the linear relationship between the independent variables (explanatory variables) and the dependent variables (response variables). Therefore, the L*, a*, S, I, and R were used as independent variables while the TBA and TVB-N were used as dependent variables, respectively, while in order to compare the influence of changes in the independent variables on the model coefficients of determination, three corresponding multiple linear regression models were established separately. The first group took the color parameters R, a*, I and S as the independent variable. The color parameters R, L*, a*, S and R, S, I, L*, a* were used as independent variables for the second and third groups, respectively. Variance inflation factor (VIF) was utilized as a measurement to evaluate the multi-collinearity among the independent variables in the multiple regression model (Liu et al., 2008). When the VIF values of the respective variables showed less than 10, it indicated that there was no significant multi-collinearity and the regression model was well fitted (Yoon and Lee, 2021).

The deterioration degree of smoked chicken thighs cannot be determined directly by color change during storage at 4 °C. Therefore, developing a fast and intuitive method to evaluate the freshness and quality of smoked chicken thighs was necessary. The effective approach to obtaining the freshness of smoked chicken thighs during storage was visualizing TVB-N and TBA values. There were corresponding color data for each pixel of the image. Hence, the freshness metrics of each pixel can be predicted through multiplying each pixel of the image by the regression coefficient of the model and visualized it according to a linear color scale (Cheng et al., 2016). In this research, the best models will be applied to convert each pixel of the image into the corresponding color separately and finally to generate the distribution maps of TBA and TVB-N. The variation of the color in the distribution map from blue to red indicated the TVB-N and TBA values increased from low to high. Changes in TBA and TVB-N in smoked chicken thigh samples can be easily estimated by observing the colors in the distribution map.

The python programming language in the OpenCV platform (Open-Source Computer Vision Library) was utilized to carry out the image processing algorithms, multiple regression algorithms, and visualization procedures, and then uploaded to the device. Experimental data on color, TBA, TVB-N were statistically analyzed using IBM SPSS Statistics 26 (SPSS Inc., Chicago, IL, USA) software to analyze variance (ANOVA). All the results were expressed as mean ± standard deviation (M ± SD). Multiple comparisons of means were established by Duncan’s multiple range test. Differences were significantly at p < 0.05. Figures were plotted by applying Origin 2021 software (OriginLab Corp, Northampton, USA). The R² was employed to estimate the fit degree of the model.