Non-Destructive Monitoring of Sweet Pepper Samples After Selected Periods of Lacto-Fermentation

Abstract

Fermented food is characterized by positive health-promoting properties. The objective of this study was to distinguish and assess the changes in the flesh structure of sweet bell pepper samples after specific periods of fermentation in a non-destructive manner. Two cultivars of pepper, red and yellow, were subjected to lacto-fermentation. The experiments lasted 56 days and the samples were taken for analysis at the beginning of the study (0 days) and after 3, 7, 10, 14, 21, 28, and 56 days. The fermentation process was monitored based on image features, which were used to develop machine learning models distinguishing samples before and after various periods of lacto-fermentation (0, 3, 7, 10, 14, 21, 28, and 56 days). The average accuracy of the classification of red bell pepper samples was up to 93% for the model built using IBk (Lazy group). The yellow bell pepper samples were distinguished up to 90% accuracy by the LMT algorithm (Trees group). The performed study allowed us to determine the changes in pepper flesh in terms of image textures during lacto-fermentation.

1. Introduction

The sweet pepper (Capsicum annuum L.), also known as the bell pepper, is one of the most frequently cultivated vegetables in the world. Pepper is third after tomato and potato among the cultivated plants from the Solanaceae family [1]. This plant originated from regions with a warm climate and it is known for its specific environmental requirements [2,3,4]. The enormous popularity of this vegetable is determined by its specific taste, dietary, and health-promoting values [5,6]. It is an excellent source of many important bioactive compounds such as vitamins C and E, polyphenols (flavonoids, phenolic acids, rutin, luteolin, and anthocyanins), and carotenoids (lutein, β-carotene, capsanthin, capsorubin, and lycopene). The types and levels of these compounds depend on the pepper cultivar. Moreover, pepper fruit is a good source of folic acid, minerals (particularly potassium, manganese, iron, and magnesium), and dietary fiber [7,8,9,10].

Among the traditional food production methods, the fermentation process is performed by lactic acid bacteria, which convert monosaccharides and disaccharides into lactic acid and other compounds using enzymes [11]. In addition, lactic acid bacteria have antimicrobial activity, which can extend the shelf life of products. These properties are possessed by bacteria from the Lactobacillus group, which have GRAS status, i.e., they are Generally Recognized as Safe [12,13,14]. Fermented products are made from high-quality raw materials, do not contain preservatives or other artificial additives, and are characterized by exceptional taste [14]. Fermented foods are also an alternative to highly processed foods. The positive health-promoting properties of fermented vegetables are attributed to the bioactive compounds that are formed during the fermentation process. These include, among others: ACE (angiotensin-converting enzyme), lactic acid, β-casomorphins, casein phosphopeptides found in fermented dairy products, kefiran found in kefir, GABA (gamma-amino-butyric acid), and SCFA (short-chain fatty acid) [14,15].

In recent years, there has been a steady increase in demand for fermented plant products, which have been recognized as functional products that benefit human health in many ways [16]. More than 2300 pepper cultivars are registered on the European market, varying in size, color, shape, spiciness, juiciness, or flesh thickness. Commercial cultivars of pepper are available with different skin colors—red, green, orange, yellow, or purple. The differences in fruit color are due to the ability to synthesize chlorophyll or carotenoids and the degree of maturity of the fruit [17].

According to Blanco-Ríos [18], the nutritional value of peppers depends, among other factors, on the color of the fruit (cultivar), growing conditions, and post-harvest treatment. The active substances found in pepper fruits have antioxidant (high content of vitamins, mainly A, C, and E, as well as minerals and polyphenols), anti-inflammatory, anti-diabetic, antimicrobial, and immunomodulation effects. Bell pepper fruit consists of 80–90% water and therefore has a relatively short shelf life compared, for example, to root vegetables. Shortly after harvesting, the fruit wilts and quickly loses its commercial value [19]. The processing of pepper fruit, especially the use of a lactic fermentation process with lactic acid bacteria, contributes to a change in taste, texture, and aroma [20].

The quality of processed agricultural food can be monitored using computer vision systems. These systems are important for the accurate and consistent inspection of external quality parameters [21]. For example, the effect of 1 week and 3 months of lacto-fermentation was assessed on zucchini flesh, obtaining overall accuracies of up to 99.33% (IBk). Additionally, it was found that the fresh sample was completely different from lacto-fermented samples in terms of image texture [22]. For the carrot, models built based on image texture parameters provided 100% accuracy in distinguishing fresh carrot slices and samples lacto-fermented for 6 months in the case of the Multilayer Perceptron machine learning algorithm [23]. Moreover, artificial intelligence and computer vision were used to investigate the fermentation level of cocoa beans in the work of Anggraini et al. [24], who found an accuracy of 94% (Multilayer Perceptron) for distinguishing fermented and unfermented cocoa and concluded that the applied approach could predict cocoa bean fermentation rate. Oliveira et al. [25] distinguished fully fermented, partially fermented, under-fermented, and unfermented cocoa beans (cut test) using computer vision, obtaining 0.93 accuracy. Machine learning algorithms based on image features were used by Bhargava et al. [26] for the grading of tea samples, which were graded as fermented, over-fermented, and under-fermented. The algorithms detected the tea fermentation level with an accuracy reaching 98.75% (SVM), 89.72% (SRC), and 87.39% (k-NN). In the case of black tea, machine learning algorithms were also applied for quality assurance and to detect the degree of fermentation based on electrical properties by Zhu et al. [27], resulting in an accuracy of up to 100% (Random Forest). The reported results indicated the usefulness of image features and artificial intelligence for the classification of fermented products with very high accuracy reaching 100%. The above literature data provided a perspective for carrying out research aimed at the non-destructive evaluation of the quality of food products using image analysis and machine learning and justified the current work. It was assumed that this approach could be used for non-destructive and objective monitoring of pepper samples during fermentation.

In this context, this study aimed at non-destructively distinguishing bell pepper samples after specific periods of lacto-fermentation to assess the changes that occur during the process. Peppers with different fruit colors can have different properties and, consequently, processing, including fermentation, may take different courses. Therefore, pepper raw materials with completely different colors were selected. The identification of the pepper fermentation step is important and can be necessary to determine after what period the greatest changes in the structure of the flesh of fermented samples occur and when the changes are the smallest. Based on this information, the conditions and duration of fermentation can be established to obtain the most desired product.

2. Materials and Methods

2.1. Fresh and Lacto-Fermented Yellow and Red Pepper Samples

The research material consisted of yellow bell pepper ‘Yellow California’ and red bell pepper ‘California’ (Figure 1). Samples were obtained from a Polish company that imports and distributes fruit and vegetables.

The pepper fruit samples without signs of disease, softening, or mechanical damage were washed thoroughly with water, then cut into four parts, and the seeds were removed. The pepper pieces were put into sterile 1000 mL glass jars with standard spices for fermentation (garlic, mustard seeds, black pepper, bay leaves, Jamaica pepper—the amount of spices was up to 2% of the quantity of pepper in the jar). Then, the jars were filled with tap water with added salt so that the final NaCl content in the brine was 3.5%. The closed jars were kept in an air-conditioned room at 20 °C to perform the spontaneous fermentation. The experiment was carried out in three technological repetitions per cultivar per one period of analysis, with three jars of fermented peppers in each repetition. The differences between raw material and samples after 3, 7, 10, 14, 21, 28, and 56 days of lacto-fermentation were analyzed non-destructively using image analysis and machine learning, as presented in Figure 2.

2.2. Image Acquisition and Processing

The color image acquisition of red and yellow sweet bell pepper samples (pieces), both in fresh and lacto-fermented forms, was performed with the use of a digital camera with F 2.4, 8× digital zoom, auto white balance, and optical image stabilization, and LED (light emitting diodes) illumination. For the lacto-fermented samples, excess liquid was removed with a paper towel. The imaging was performed on a black background. The image acquisition was performed in one hundred repetitions. For each pepper class, ten images with ten pieces in each image were obtained. During the image processing, each image was cropped into ten images including one pepper piece each. Therefore, in total, one hundred images were obtained for each group. The pepper sample images are shown in Figure 3. The images were processed using the Mazda 4.7 software (Łódź University of Technology, Institute of Electronics, Łódź, Poland) [28,29,30]. The images were segmented to separate pepper samples from the background, select regions of interest (ROIs) as pepper pieces, and compute image features. For each pepper sample, 2172 texture parameters were extracted from images in different color channels L, a, b, R, G, B, X, Y, Z, S, V, and U.

2.3. Classification Model Development

Models for change determination in the structure of yellow and red bell pepper samples occurring during the lacto-fermentation were built using machine learning algorithms by WEKA 3.9 software (Waikato Environment for Knowledge Analysis, University of Waikato, Hamilton, New Zealand) [31,32,33]. For the development of models, selected attributes with the highest discriminative power were used. Image texture selection was performed using the Best First search algorithm. The pepper samples after specific periods of fermentation were classified using selected machine learning algorithms from groups of Functions, Bayes, Meta, Lazy, Trees, and Rules. The test mode of 10-fold cross-validation was applied. This means that separate datasets were used for validation than those for training. Each dataset included 800 cases (100 cases for each of the 8 pepper classes). The datasets were randomly divided into 10 parts. Nine parts were considered as the training sets and one part was the validation set. The process was repeated 10 times with different validation sets and the overall results were the averages of 10 estimations. Ten folds ensured high results for classification metrics and reproducibility of results. As classification performance metrics, accuracies, True-Positive Rate (TP Rate), False-Positive Rate (FP Rate), Precision, Recall, F-Measure, Matthews Correlation Coefficient (MCC), F-Measure, Precision–Recall Area (PRC Area), Receiver Operating Characteristic Area (ROC Area), and Kappa statistic were computed. Equations (1)–(10) were used [34,35,36,37].

$$\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}=\mathrm{}{\displaystyle \frac{(\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N})}{\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{N}+\mathrm{F}\mathrm{P}}}$$

(1)

$$\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}=\mathrm{}{\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}}}$$

(2)

$$\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}=\mathrm{}{\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}}$$

(3)

$$\mathrm{F}-\mathrm{m}\mathrm{e}\mathrm{a}\mathrm{s}\mathrm{u}\mathrm{r}\mathrm{e}=\mathrm{}{\displaystyle \frac{2 \mathrm{*} \mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n} \mathrm{*} \mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}{(\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}+\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l})}}$$

(4)

$$\mathrm{M}\mathrm{C}\mathrm{C}=\mathrm{}{\displaystyle \frac{\left(\mathrm{T}\mathrm{P} \mathrm{*} \mathrm{T}\mathrm{N}-\mathrm{F}\mathrm{P} \mathrm{*} \mathrm{F}\mathrm{N}\right)}{\sqrt{\left(\left(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}\right)\left(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}\right)\left(\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{P}\right)\left(\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{N}\right)\right)}}}$$

(5)

$$\mathrm{T}\mathrm{P}\mathrm{R}\mathrm{}=\mathrm{}{\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}}$$

(6)

$$\mathrm{F}\mathrm{P}\mathrm{R}=\mathrm{}{\displaystyle \frac{\mathrm{F}\mathrm{P}}{\mathrm{F}\mathrm{P}\mathrm{}+\mathrm{T}\mathrm{N}}}$$

(7)

$$\mathrm{R}\mathrm{O}\mathrm{C} \mathrm{A}\mathrm{r}\mathrm{e}\mathrm{a}=\mathrm{a}\mathrm{r}\mathrm{e}\mathrm{a} \mathrm{u}\mathrm{n}\mathrm{d}\mathrm{e}\mathrm{r} \mathrm{T}\mathrm{P}\mathrm{R} \mathrm{v}\mathrm{s}. \mathrm{F}\mathrm{P}\mathrm{R} \mathrm{c}\mathrm{u}\mathrm{r}\mathrm{v}\mathrm{e}$$

(8)

$$\mathrm{P}\mathrm{R}\mathrm{C} \mathrm{A}\mathrm{r}\mathrm{e}\mathrm{a}=\mathrm{a}\mathrm{r}\mathrm{e}\mathrm{a} \mathrm{u}\mathrm{n}\mathrm{d}\mathrm{e}\mathrm{r} \mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n} \mathrm{v}\mathrm{s}. \mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l} \mathrm{c}\mathrm{u}\mathrm{r}\mathrm{v}\mathrm{e}$$

(9)

$$\mathrm{K}\mathrm{a}\mathrm{p}\mathrm{p}\mathrm{a}=\mathrm{}{\displaystyle \frac{\frac{\left(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}\right)\left(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}\right)}{\left(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}\right)\left(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}\right)\left(\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{P}\right)\left(\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{N}\right)}\mathrm{}+\mathrm{}\frac{\left(\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{P}\right)\left(\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{N}\right)}{\left(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}\right)\left(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}\right)\left(\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{P}\right)\left(\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{N}\right)}\mathrm{}}{\left(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}\right)\left(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}\right)\left(\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{P}\right)\left(\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{N}\right)}}$$

(10)

where TP is True Positive; TN is True Negative; FP is False Positive; FN is False Negative; TPR is the True-Positive Rate; and FPR is the False-Positive Rate.

The choice of the models used for the classification of red bell pepper samples before and after various periods of lacto-fermentation was justified by providing the highest accuracies and values of the other above-mentioned metrics. The most successful models for distinguishing red bell pepper samples were models built using IBk, WiSARD, and Random Committee algorithms, whereas in the case of the classification of yellow bell pepper samples, LMT and IBk algorithms provided the highest results.

3. Results

3.1. The Classification of Red Bell Pepper Samples Before and After Various Periods of Lacto-Fermentation

The red bell pepper samples before and after various periods of lacto-fermentation (0, 3, 7, 10, 14, 21, 28, and 56 days) were distinguished using machine learning algorithms. The following image texture parameters of red bell pepper were characterized by the highest discriminative power: RSGArea, RS5SH1SumVarnc, RS5SH1DifEntrp, RS5SH3SumVarnc, RS5SZ3SumVarnc, RATeta1, GHPerc01, BS4RNGLevNonU, LHPerc01, XHPerc01, XSGNonZeros, XSGPerc01, XS5SH1SumVarnc, XS5SZ1AngScMom, UHPerc10, and UHDomn01. The average accuracy of the classification reached 93% for the model built using IBk from the Lazy group (

Table 1. The confusion matrix of the classification of red bell pepper samples before and after various periods of lacto-fermentation using the IBk algorithm.

Algorithm	Predicted Class (Days of Fermentation) (%)								Actual Class (Days of Fermentation)	Average Accuracy (%)
	0 Days	3 Days	7 Days	10 Days	14 Days	21 Days	28 Days	56 Days
IBk	100	0	0	0	0	0	0	0	0 days	93
	0	96	0	0	0	0	0	4	3 days
	0	0	100	0	0	0	0	0	7 days
	0	4	0	72	8	8	4	4	10 days
	0	0	0	4	86	4	0	6	14 days
	0	0	0	3	0	97	0	0	21 days
	0	0	0	4	4	0	89	3	28 days
	0	0	0	0	0	0	0	100	56 days

Blue—correctly classified cases; green—incorrectly classified cases.

). The lowest classification accuracy (72%) was obtained for a red bell pepper sample after 10 days of the process. The greatest number of misclassified cases occurred between samples fermented for 10, 14, and 21 days. As many as 8% of cases belonging to the actual class of samples after 10 days of fermentation were incorrectly classified as samples after 14 days of fermentation, and 8% of cases were incorrectly classified as pepper fruit after 21 days of the process. In the case of the model developed using the IBk algorithm, the time taken to build the model was 0.01 s and the Kappa statistic was equal to 0.9206.

The detailed performance metrics presented in

Table 2. Detailed classification performance metrics of fresh and lacto-fermented red bell pepper samples (IBk algorithm).

Algorithm	Class (Days of Fermentation)	TP Rate	FP Rate	Precision	Recall	F-Measure	MCC	ROC Area	PRC Area
IBk	0 days	1.000	0.000	1.000	1.000	1.000	1.000	1.000	1.000
	3 days	0.958	0.005	0.958	0.958	0.958	0.954	0.974	0.919
	7 days	1.000	0.000	1.000	1.000	1.000	1.000	1.000	1.000
	10 days	0.720	0.015	0.857	0.720	0.783	0.762	0.880	0.702
	14 days	0.857	0.015	0.889	0.857	0.873	0.856	0.933	0.815
	21 days	0.968	0.015	0.909	0.968	0.937	0.928	0.974	0.876
	28 days	0.893	0.005	0.962	0.893	0.926	0.917	0.941	0.902
	56 days	1.000	0.025	0.868	1.000	0.930	0.920	0.986	0.855

TP Rate (True-Positive Rate), FP Rate (False-Positive Rate), MCC (Matthews Correlation Coefficient), PRC Area (Precision–Recall Area), and ROC Area (Receiver Operating Characteristic Area).

confirmed the highest classification accuracy of unfermented pepper and the sample after 7 days of lacto-fermentation. It was expressed by the lower value of FT Rate equal to 0.000 and the highest values of all remaining parameters, such as TP Rate, Precision, Recall, F-Measure, MCC, F-Measure, PRC Area, and ROC Area equal to 1.000. The lowest values of TP Rate, Precision, Recall, F-Measure, MCC, F-Measure, PRC Area, and ROC Area were found for samples after 10 and 14 days.

The model built by the WiSARD algorithm from the group of Functions allowed for the classification of red bell pepper samples with an average accuracy of 92% (

Table 3. The classification accuracies of fresh and lacto-fermented red bell pepper using the WiSARD algorithm.

Algorithm	Predicted Class (Days of Fermentation) (%)								Actual Class (Days of Fermentation)	Average Accuracy (%)
	0 Days	3 Days	7 Days	10 Days	14 Days	21 Days	28 Days	56 Days
WiSARD	100	0	0	0	0	0	0	0	0 days	92
	0	100	0	0	0	0	0	0	3 days
	0	0	100	0	0	0	0	0	7 days
	4	4	0	72	8	4	0	8	10 days
	0	0	0	4	82	7	0	7	14 days
	0	0	0	10	0	90	0	0	21 days
	0	0	0	4	0	0	93	3	28 days
	0	0	0	0	0	0	0	100	56 days

Blue—correctly classified cases; green—incorrectly classified cases.

). Pepper samples analyzed on days 0, 3, and 7 were characterized by an accuracy of 100%. An accuracy of 100% was also found for the sample after 56 days of the process, whereas the samples that underwent 10 days of fermentation were classified with the lowest accuracy of 72%. The model development took 0.08 s. The value of the Kappa statistic was very high: equal to 0.9157.

The samples at 0, 3, 7, and 56 days of fermentation were characterized by TP Rate and Recall of 1.000 (

Table 4. Classification performance metrics of fresh and lacto-fermented red bell pepper samples (WiSARD algorithm).

Algorithm	Class (Days of Fermentation)	TP Rate	FP Rate	Precision	Recal	F-Measure	MCC	ROC Area	PRC Area
WiSARD	0 days	1.000	0.005	0.970	1.000	0.985	0.982	0.987	0.951
	3 days	1.000	0.005	0.960	1.000	0.980	0.977	0.912	0.825
	7 days	1.000	0.000	1.000	1.000	1.000	1.000	1.000	1.000
	10 days	0.720	0.024	0.783	0.720	0.750	0.722	0.797	0.522
	14 days	0.821	0.010	0.920	0.821	0.868	0.853	0.955	0.817
	21 days	0.903	0.015	0.903	0.903	0.903	0.888	0.933	0.823
	28 days	0.929	0.000	1.000	0.929	0.963	0.959	0.900	0.828
	56 days	1.000	0.025	0.868	1.000	0.930	0.920	0.964	0.846

TP Rate (True-Positive Rate), FP Rate (False-Positive Rate), MCC (Matthews Correlation Coefficient), PRC Area (Precision–Recall Area), and ROC Area (Receiver Operating Characteristic Area).

). It confirmed that the accuracy of classification was equal to 100%. The lowest values of most of the analyzed metrics were observed for samples after 10, 14, and 21 days.

The classification model developed using the Random Committee algorithm from the Meta group correctly classified pepper samples after various periods of fermentation in 88% of cases (

Table 5. The confusion matrix of distinguishing red bell pepper samples before and after lacto-fermentation for different periods using the Random Committee algorithm.

Algorithm	Predicted Class (Days of Fermentation) (%)								Actual Class (Days of Fermentation)	Average Accuracy (%)
	0 Days	3 Days	7 Days	10 Days	14 Days	21 Days	28 Days	56 Days
Random Committee	97	0	0	0	0	0	0	3	0 days	88
	4	88	0	0	0	0	0	8	3 days
	0	3	97	0	0	0	0	0	7 days
	4	0	0	72	12	4	4	4	10 days
	0	0	0	11	75	7	0	7	14 days
	0	0	0	0	3	94	3	0	21 days
	0	0	0	0	4	0	96	0	28 days
	0	3	0	3	3	3	0	88	56 days

Blue—correctly classified cases; green—incorrectly classified cases.

). The pepper sample after 10 days of fermentation was classified with the lowest correctness of 72%. The highest number of misclassified cases occurred between samples fermented for 10 and 14 days. This indicated the highest similarity of these samples in terms of structure. As many as 12% of cases from the peppers fermented for 10 days were incorrectly classified as peppers that had been fermented for 14 days and 11% of cases from the actual class of pepper fermented for 14 days were misclassified as fruit fermented for 10 days. The model was built in 0.02 s. The Kappa statistic was 0.8711.

The Random Committee algorithm allowed for slightly lower classification correctness. None of the samples reached a value of 1.000 for the metrics, such as TP Rate, Recall, F-Measure, MCC, and PRC Area (

Table 6. Detailed classification performance metrics of red bell pepper samples before and after various periods of lacto-fermentation using the Random Committee algorithm.

Algorithm	Class (Days of Fermentation)	TP Rate	FP Rate	Precision	Recall	F-Measure	MCC	ROC Area	PRC Area
Random Committee	0 days	0.969	0.010	0.939	0.969	0.954	0.946	0.997	0.978
	3 days	0.875	0.010	0.913	0.875	0.894	0.882	0.987	0.925
	7 days	0.967	0.000	1.000	0.967	0.983	0.981	1.000	0.999
	10 days	0.720	0.019	0.818	0.720	0.766	0.741	0.896	0.718
	14 days	0.750	0.030	0.778	0.750	0.764	0.732	0.943	0.859
	21 days	0.935	0.020	0.879	0.935	0.906	0.892	0.983	0.888
	28 days	0.964	0.010	0.931	0.964	0.947	0.940	0.997	0.977
	56 days	0.879	0.030	0.829	0.879	0.853	0.828	0.956	0.853

TP Rate (True-Positive Rate), FP Rate (False-Positive Rate), MCC (Matthews Correlation Coefficient), PRC Area (Precision–Recall Area), and ROC Area (Receiver Operating Characteristic Area).

). The sample fermented for 7 days was characterized by Precision and ROC Area equal to 1.000 and FP Rate of 0.000. The classification performance metrics showed the lowest correctness of distinguishing samples after 10 and 14 days of the process.

3.2. The Classification of Yellow Bell Pepper Samples Before and After Various Periods of Lacto-Fermentation

In the case of the yellow bell pepper, the image textures with the highest discriminative power, which were the most important for building the most successful models, were RHMaxm01, RS5SV1SumVarnc, RS4RVLngREmph, GS5SZ5DifVarnc, LHSkewness, LHKurtosis, aHSkewness, aHKurtosis, aHDomn10, XHSkewness, YHMean, ZS5SZ3DifEntrp, US4RVGLevNonU, UASigma, SHPerc99, SS5SH5InvDfMom, SS5SN5SumVarnc, and SASigma. The samples of yellow bell pepper subjected to lacto-fermentation for various periods were also distinguished with high accuracy using machine learning models built based on selected image features. The LMT algorithm from the Trees group allowed for the classification of samples on days 0, 3, 7, 10, 14, 21, 28, and 56, with an average accuracy of 90% (

Table 7. The classification of yellow bell pepper samples before and after different periods of lacto-fermentation using the LMT algorithm.

Algorithm	Predicted Class (Days of Fermentation) (%)								Actual class (Days of Fermentation)	Average Accuracy (%)
	0 Days	3 Days	7 Days	10 Days	14 Days	21 Days	28 Days	56 Days
LMT	100	0	0	0	0	0	0	0	0 days	90
	0	87	0	0	3	3	3	3	3 days
	0	0	96	0	0	0	0	4	7 days
	0	0	0	76	16	5	3	0	10 days
	0	0	0	8	85	0	0	8	14 days
	0	0	0	0	3	90	3	3	21 days
	0	0	0	0	6	0	88	6	28 days
	0	0	0	2	0	0	0	98	56 days

Blue—correctly classified cases; green—incorrectly classified cases.

). The greatest changes in the pepper structure occurred at the beginning of the fermentation. The unfermented sample was completely different (100% accuracy) from all samples after fermentation. The greatest number of misclassified cases occurred between the 10th day and 14th day. The time taken to build a model using LMT was 1.31 s and the Kappa statistic of 0.8956 was determined.

The value of FT Rate equal to 0.000 and the values of TP Rate, Precision, Recall, F-Measure, MCC, F-Measure, PRC Area, and ROC Area equal to 1.000 for the yellow bell pepper sample before fermentation indicated its complete difference from the fermented samples in terms of image textures (

Table 8. Detailed metrics of the classification of fresh and lacto-fermented yellow bell pepper using the LMT algorithm.

Algorithm	Class (Days of Fermentation)	TP Rate	FP Rate	Precision	Recall	F-Measure	MCC	ROC Area	PRC Area
LMT	0 days	1.000	0.000	1.000	1.000	1.000	1.000	1.000	1.000
	3 days	0.867	0.000	1.000	0.867	0.929	0.924	0.946	0.916
	7 days	0.963	0.000	1.000	0.963	0.981	0.979	0.976	0.968
	10 days	0.757	0.012	0.903	0.757	0.824	0.804	0.927	0.834
	14 days	0.846	0.038	0.688	0.846	0.759	0.737	0.975	0.644
	21 days	0.900	0.012	0.900	0.900	0.900	0.888	0.989	0.936
	28 days	0.879	0.012	0.906	0.879	0.892	0.879	0.993	0.947
	56 days	0.976	0.028	0.854	0.976	0.911	0.898	0.997	0.983

TP Rate (True-Positive Rate), FP Rate (False-Positive Rate), MCC (Matthews Correlation Coefficient), PRC Area (Precision–Recall Area), and ROC Area (Receiver Operating Characteristic Area).

).

A slightly lower average accuracy of the classification (89%) of yellow bell pepper samples after various periods of fermentation was obtained for a model built using the IBk algorithm from the Lazy group (

Table 9. The accuracies of the classification of fresh and lacto-fermented yellow bell pepper samples using the IBk algorithm.

Algorithm	Predicted Class (Days of fermentation) (%)								Actual Class (Days of Fermentation)	Average Accuracy (%)
	0 Days	3 Days	7 Days	10 Days	14 Days	21 Days	28 Days	56 Days
IBk	95	0	0	0	0	2	0	3	0 days	89
	0	90	0	3	3	0	3	0	3 days
	7	0	89	0	0	4	0	0	7 days
	3	0	0	76	19	3	0	0	10 days
	0	0	0	31	65	0	0	4	14 days
	0	0	0	0	0	100	0	0	21 days
	0	3	0	0	0	0	97	0	28 days
	0	0	0	0	0	0	0	100	56 days

Blue—correctly classified cases; green—incorrectly classified cases.

). The lowest accuracies were observed for fermented pepper samples for 10 and 14 days: 76 and 65%, respectively. The highest number of misclassified cases between these samples indicated the greatest similarity between them. The model was developed in 0.01 s and the Kappa statistic reached 0.8875.

The high classification accuracies of samples fermented for 21 and 56 days were confirmed by the highest values of TP Rate and Recall, whereas the lowest values of these parameters were observed for pepper samples after 10 and 14 days of fermentation (

Table 10. Classification performance metrics of yellow bell pepper samples before and after various periods of lacto-fermentation (IBk algorithm).

Algorithm	Class (Days of Fermentation)	TP Rate	FP Rate	Precision	Recall	F-Measure	MCC	ROC Area	PRC Area
IBk	0 days	0.953	0.013	0.953	0.953	0.953	0.940	0.976	0.926
	3 days	0.900	0.004	0.964	0.900	0.931	0.924	0.934	0.853
	7 days	0.889	0.000	1.000	0.889	0.941	0.937	0.956	0.902
	10 days	0.757	0.036	0.757	0.757	0.757	0.721	0.888	0.662
	14 days	0.654	0.030	0.680	0.654	0.667	0.635	0.875	0.551
	21 days	1.000	0.012	0.909	1.000	0.952	0.948	0.995	0.918
	28 days	0.970	0.004	0.970	0.970	0.970	0.966	0.968	0.947
	56 days	1.000	0.012	0.933	1.000	0.966	0.960	0.992	0.921

TP Rate (True-Positive Rate), FP Rate (False-Positive Rate), MCC (Matthews Correlation Coefficient), PRC Area (Precision–Recall Area), and ROC Area (Receiver Operating Characteristic Area).

).

4. Discussion

In our study, the application of image analysis of pepper samples and machine learning models developed based on texture image parameters allowed for the non-destructive and objective determination of differences in samples before and after selected periods of lacto-fermentation. The image texture features were used to distinguish the samples. The image texture is a function of the spatial variation of the pixel brightness intensity. Image textures can provide important information about the structure of the evaluated samples. The quantitative analyses of texture parameters from pictures provide valuable insights into product quality [28,38]. Therefore, the changes in the quality of peppers during the fermentation were evaluated based on pictures in terms of flesh structure.

The determined correctness of classification for individual samples revealed the changes in the structure of the peppers during the process. It was observed that the classification accuracy depended on the pepper cultivar. In the case of red bell pepper, the average accuracy of distinguishing samples before and after 3, 7, 10, 14, 21, 28, and 56 days of lacto-fermentation reached 93%, whereas the samples of yellow bell pepper were correctly classified with an average accuracy of up to 90%. This meant that the differences between samples were slightly greater for red pepper. Also, Janiszewska-Turak et al. [17] reported that the properties of fermented pepper depend on the cultivar.

Furthermore, the accuracy of distinguishing pepper samples depended on the machine learning algorithm used to build the models. Among the many tested algorithms from the Functions, Bayes, Meta, Lazy, Trees, and Rules groups, the highest results were obtained for IBk (Lazy)—93%, WiSARD (Functions)—92%, and Random Committee (Meta)—88% in the case of red bell pepper and LMT (Trees)—90% and IBk (Lazy)—89% for yellow bell pepper samples. Also, the literature data reported the usefulness of image analysis and machine learning for the quality assessment of fermented pepper. In the previous study regarding red bell pepper, the use of IBk, SMO, Random Forest, Naive Bayes, Filtered Classifier, and JRip machine learning algorithms allowed for fresh and 6-month lacto-fermented samples to be distinguished with an average accuracy of 99% [39].

The obtained results can be of practical application. A vision system that takes into account the use of imaging devices can be used to determine the correct course of the fermentation process. Additionally, knowing the course of fermentation for different cultivars and texture parameters of the flesh of raw pepper can be useful for planning the process. However, the applied approach also has limitations. The practical problem of fermentation is the fact that the course of the process and the quality of fermented products depend on the cultivar, degree of maturity, and physicochemical properties of the raw material [40,41]. To scale up the laboratory experiments to practical applications, it is necessary to collect more data for a larger number of cultivars, degrees of maturity, and growing seasons.

Although spontaneous fermentation is a reliable way of preserving fruit and vegetables, there is a risk of fermentation failure, undesirable and unpredictable changes in the quality properties, or insufficient inhibition of microorganisms causing product spoilage [42]. The application of spontaneous/natural fermentation has also disadvantages related to the fact that spoilage organisms can be easily introduced, fermentation strains are unknown, starting fermentation is difficult, and the results of the process are not controllable [43]. Therefore, in our study, slight differences could have occurred between individual technological repetitions, influencing the properties of the products and the accuracies of their classification. To avoid the problem of uncontrolled fermentation and ensure greater safety and higher quality of fermented products, selected starter cultures can be used to perform the process [44].

5. Conclusions

The obtained results revealed that image analysis combined with machine learning models allowed for the non-destructive monitoring of differences in pepper samples after selected periods of lacto-fermentation. The changes in image textures reflected changes in the flesh structure of the pepper samples. Determining the correctness of the classification of samples on days 0, 3, 7, 10, 14, 21, 28, and 56, it was found that the accuracy depended on the pepper cultivar (red or yellow) and the machine learning algorithm applied to build the models. The differences between samples were greater for red than yellow bell peppers. The highest classification results were obtained for the IBk algorithm for red pepper and LMT for yellow pepper. Further research can focus on monitoring the changes in pepper samples using more cultivars, different degrees of maturity, and raw materials from different growing seasons. Additionally, the comparison of the effect of spontaneous fermentation and the process with the use of starter cultures on the structure of pepper flesh can be carried out.