Prediction of food quality parameters in fish burgers by partial least square models using RGB pattern of digital images

Abstract

Abstract

Rancid taste, pH, and TBARS are important quality parameters of food oxidation, analyzed in a time-consuming and destructive way. Non-destructive characterization of food can be achieved correlating this data with computational vision. Thus, the present study aimed to use RGB digital images to predict sensory rancid taste, pH, and TBARS results in fish burgers. A mobile obtained the digital images, in a controlled environment, and 768 grayscales were performed using RGB histograms. The pH, showed a peak at 21st day of storage, which PCA confirmed by isolating the 21st samples, corroborated by HCA grouping 21st day samples. PLS models from RGB digital images and sensory rancidity, pH and TBARS data, using mean center method and SIMPLS algorithm found models with > 0.97 R². Thus, any digital image of this batch of burgers, inserted into the model to predict rancid taste, pH and TBARS has high confidence level of prediction.

Graphical abstract

Introduction

Digital images are two-dimensional arrays containing millions of squares known as pixels. Also are matrices with M lines (height of the image) and N columns (width of the image), according to the color model applied. In the RGB, color model each color tone is defined by three channels: R (red), G (green) and B (blue), which vary from 0 (black) to 255 (white) (Damasceno et al. 2015; Helfer et al. 2016). Color is directly related to food content, biscuits for examples have dark and light tones of the yellow colour and meat which contains fat have white tones (Tarak et al., 2016). RGB models are being widely studied due to its advantages over CIE/Hunter L*a*b*, HVS and CMYK models: Colour intensity, saturation and brightness offered by RGB are much superior compared to others, the reason that RGB is well supported by many image processing programs; RGB patterns can be converted into CMYK and L*a*b* and offers bright color before conversion (Manamohana et al. 2020).

Several devices are able to capture digital images as scanners, cellphones, and digital cameras. When images are decomposed into a color diagram, they can be treated in the same way as spectrophotometric measurements by multivariate calibration (as RPCA – Regression by Principal Components Analysis; and PLS – Partial Least Squares) (Helfer et al. 2016; Botelho et al. 2017; de Camargo et al. 2017).

Conventional spectroscopic and chemical methods are often sampling destructive, time-consuming, expensive and it might need specialized workers. Therefore, image analysis has been considered as a highly suitable and non-destructive alternate method, used for real-time and field identification of chemical compounds without wastes (Dowlati et al. 2013; Kucheryavskiy et al. 2014; Damasceno et al. 2015; Benalia et al. 2016; Zapotoczny et al. 2016).

The literature meets many digital imaging studies these days, in whole fish (Dowlati et al. 2013), fish fillets (Mateo et al. 2006), fish freshness (Maravillas et al. 2018; Lalabadi et al. 2020), meat (Sun et al. 2016; Zapotoczny et al. 2016; Cruz-Fernandez et al. 2017). Moreover, food in general, as milk (Kucheryavskiy et al. 2014), honey (Dominguez and Centurión 2015), drinking water (Damasceno et al. 2015), chocolate (Briones and Aguilera 2005), candies (Botelho et al. 2017), banana (Intaravanne et al. 2012), etc.

Freshness is one of the main parameters important to the quality of fish and its products and it can be accessed by chemical methods as pH and rancid evaluations like TBARS. The fat from fish is highly susceptible to rancidity due to the content of the unsaturated omega-type. Research showed no studies of digital imaging to determine quality descriptors of fish burgers, in order to obtain robustly models with PLS. Thus, the main contribution of this work is to calibrate a method based on digital images to predict pH, TBARS and rancid taste for quality control of grass carp burgers employing PCA and PLS.

Materials and methods

Samples preparation

The burgers were prepared following adapted formulation from Velioğlu et al. (2010) as described by Marques et al. (2017), with grass carp fish, and the samples were stored under refrigeration during 30 days (4 °C ± 2 °C), vacuum packed.

Apparatus and software

A Moto G cellphone first generation (5 megapixels camera, 2592 × 1944 pixels, autofocus, flash LED) obtained the digital images. Chemometric data treatment and multivariate statistics (PCA, PLS – SIMPLS algorithm) were managed with ChemoStat® software (Helfer et al. 2015). The 49 most relevant grayscale used for the multivariate analysis were selected from the initial 768 by the genetic algorithm in the software Weka® Data Mining. The cluster analysis (HCA) was performed through Past® software.

pH, TBARS and chemical analyses

The pH of the burgers was measured using a bench top equipment (TECNAL®) and the thiobarbituric acid reactive substances (TBARS) were determined according to AOAC (AOAC - Association Of Official Analytical Chemistry 2000). The chemical characterization of the product was analyzed according to AOAC (2000) for moisture, ash and protein, and lipids via Bligh and Dyer (1959). The carbohydrate content was obtained by difference.

Rancid taste evaluation

A previously trained panel (Marques et al. 2017) of 7 assessors evaluated the rancid flavor. A 10 cm unstructured scale was applied to distribute the samples within this range, anchored in “no rancid flavor” and “pronounced rancid flavor”.

Digital image acquisition and analysis

The burgers were removed from the refrigerator (4  ± 2 °C), placed into a Petri dish on a blank A4 paper, with natural light and room temperature. The cellphone was placed in a fixed position in the center of the A4 paper above the Petri dish (see Graphical Abstract). The distance from the camera to the paper sheet was about 20 cm and the smartphone was held manual (Valderrama and Valderrama 2016). Six replicates of 25 g had the digital image captured each day of analysis during 30 days (Days 1, 7, 14, 17, 21, 23, 25 and 30).

768 grayscales were performed using RGB histogram from the digital images. Genetic algorithm was used to resize the data and 49 shades of gray were selected for the models. The selection of the best attributes were carried out in the Weka software.

The 48 samples were randomly assigned to calibration – 70% of samples (33) and prediction – internal validation – 30% of samples (15). The models were run using the ChemoStatV2 software. Two models with 49 or 768 grayscales were made and discussed in this work. The performance of the models were evaluated by Correlation coefficient (R² > 0.99), latent variables number, root mean square errors of calibration (RMSEC) and root mean square errors of cross validation (RMSECV) (Damasceno et al. 2015), as described by the Eqs. 1 and 2.

$$R^2=1-\frac{\text{Sum}\text{Squared}\text{Regression}\text{Error}}{\text{Sum}\text{Squared}\text{Total}\text{Error}}$$ 1

$$\text{RMSEC}\text{and}\text{RMSECV}=\sqrt{\frac{\sum _{i=1}^n\left(y_{\text{cal}}-y_{\text{exp}}\right)2}{n}}\left(2\right)$$ 2

n = samples on calibration or cross validation; $y_{\mathit{cal}}=$ the measurement predicted and $y_{\mathit{exp}}=$ is the experimental measurement.

Results and discussion

The portion of fat in the burgers analyzed by digital images was significant, 6.34 ± 0.60%, when compared to the fat in the raw material (grass carp) which was, 1.73 ± 0.32%. Accordingly, protein content was 15.12 ± 0.56%, moisture 74.79 ± 0.56%, ashes 1.87 ± 0.53% and carbohydrates 1.88 ± 0.25 in the fish burger under study.

Behavior of the input data

The pH results behavior showed an increase in the 21st day (Data shown in Marques et al. 2019) of shelf life and immediate decrease after that, indicating the production of biogenic amines (organic bases). This may have improved the pH and propitiated the possible proliferation of lactic acid psychotropic bacteria (LAB) with the production of lactic acid, which inhibited its own growth, and lowered the pH again. The reduced O₂, from vacuum packing, contributed to this behavior, but also prevented the higher extend of oxidative rancidity and aerobic microbial spoilage (Masniyom 2011). Moreover, TBARS data were adjusted to an exponential model with a coefficient of determination, R² of 0.99. Sensory results presented many zero scores for the rancid taste in the initial days (1 and 7) with exponential growth right after until the 21st day of storage, also with R² > 0.95. The peak of rancid taste corresponded with the pH peak.

Multivariate calibration results

After adding the images to the ChemoStat V2 software, the 768 grayscales were extracted, and 49 most important of those were selected by the genetic algorithm in Weka software. Through these selected grayscales of the images, the PLS was performed, in addition to the PCA and grouping by cluster (HCA). The results presented are SIMPLS algorithm using the autoscale as the data preprocessing, available in ChemoStatV2, which obtained results superior to the other preprocessing method, called the mean center.

The SIMPLS algorithm, contrary to the other PLS algorithm known as NIPALS, was designed to maximize the covariance by calculating PLS factors as linear combinations of the original variables (Martins et al. 2010). In the present findings, the values with SIMPLS algorithm were superior to the R² of the NIPALS models. The results displayed in Table 1 allow us to notice when used all the grayscales (768), a smaller number of factors is necessary to reach an R² of 0.99. With the same 18 factors to obtain 0.99 as R² in the pH model, 49 grayscales reach 0.97.

Table 1.

– Resulting parameters from the PLS models obtained for the pH, TBARS and rancid taste data for fish burgers

Variable	Grayscales	Factors	R²	RMSEC	RMSECV
pH	49	25	0.9902	0.0688	0.6020
	768	18	0.9914	0.0656	0.5920
TBARS	49	34	0.9902	0.0366	0.4830
	768	17	0.9921	0.0375	0.1927
Rancid taste	49	28	0.9928	0.3633	0.8810
	768	18	0.9914	0.3770	2.4294

Table 1.

The R² is not a variable directly proportional to the RMSEC and RMSECV values. Knowing that, the lower these values, the better is the model quality, i.e., that using all the factors has not necessarily decreased and improved both parameters. This result demonstrated that the choice of data mining (greyscales) is a useful step in the process of creating models by PLS.

The representation of the images considered outliers according to the extraction of grayscales (Fig. 1a), confirmed that two of the samples of the 21st day were considered different from the others. This fact was corroborated by the PCA – with 75% of the explained variance (Fig. 1b) which positively correlated the samples on the 21st day, separated from others, with only one sample on the 23rd day. In addition, the HCA (Fig. 1c) differentiated four samples on the 21st day from the others, with a similarity of 60% between them. The HCA presented a cophenetic correlation coefficient of 0.77 (a value that the closer to 1.00 the better), indicating good quality comparing the grouped clusters.

Fig. 1.

a Representation of outlier samples according to the analysis of digital images.b Projection of the 48 samples on the factor-plane (PC1 X PC2).c Dendrogram of the samples grouped for similarity; cophenetic correlation coefficient = 0.77

Figure 1.

Therefore, the outlier results in the 21st day of the fish burger storage, with the highest values of pH and rancid taste, were perceptible by the recorded digital images and its RGB channels. The suitable quality of the PLS models was expressed by the low values for RMSEC and RMSECV (Fig. 2), except the model for rancid taste (768 grayscales). However, the RMSECV value found greater than 2.0 can be explained by the high amount of zero values for this variable, in the initial days of storage. An agglomeration of equal values at some point of data collection impairs the quality of the attempted model.

Fig. 2.

Correlation between values measured x by the values predicted, using a PLS regression model, for a pH, b TBARS and c rancid taste

Figure 2.

Although, still there is a challenge in fitting sensory data in mathematical models, due to lack of data homogeneity (Marques et al. 2019). Thus, a selection of the most relevant channels proved to be efficient in increasing correlations and, decreasing errors in calibration and validation.

Conclusion

With this study, it was seen that RGB patterns from digital images can be widely applied to fish burgers to predict pH, TBARS, and rancid taste for quality control by an algorithm like as PLS, and ChemoStat® is suitable software for that. The present findings showed early stages of replacement of physicochemical analyses for non-destructive instrumental assessment, which can be used for other researchers and industries. Selecting the RGB channels and their grayscale with data mining may be useful to predict and evaluate the sensory parameters of fish hamburgers. The possibility to have results of quality control without destroying the sample is very attractive, using software without the cost of installation and mobile equipment of image capture with low-cost, and possibilities to access results in the field.

Funding

This study wasn’t funded.

Data Availability

If requested.

Declarations

Conflict of interest

Caroline Marques declares that she has no conflict of interest. Carlos E. B. Toazza declares that he has no conflict of interest. Carla Cristina Lise declares that she has no conflict of interest Vanderlei A. de Lima declares that he no conflict of interest. Marina Leite Mitterer-Daltoé declares that she has no conflict of interest.

Ethics approval

The sensory data was approved by Ethics Committee – CAAE number 48687815.0.0000.5547 – UTFPR, Pato Branco/PR.

Informed consent

Informed consent was obtained from all individual participants included in the study.