Introduction

Potato (Solanum tuberosum L.) is considered one of the major food crops. Potato is a very nutritive vegetable. Potato may contain about 75% water. The cooked potatoes consist of 20% carbohydrate, 1.8% fiber, 1.7% protein, vitamin C, calcium, riboflavin, niacin, other minerals, and provide 86 kcal energy. Due to the presence of antioxidants, phenolic compounds and low cholesterol, eating potatoes may protect people against cancer, cardiovascular diseases and disorders and trauma of cells [1]. Potato may be consumed after processing, in different forms, such as, for example, salads, potato soups, mashed potatoes, French fries, potato chips [2]. Due to great importance in the human diet, research regarding various aspects of potato evaluation may prove to be very significant [3].

Potatoes are grown in hundreds of countries around the world. Therefore, due to the diverse climate and soil properties in different sites, the development of new cultivars of potato was desired. Crossbreeding resulted in the occurrence of thousands of potato cultivars [4]. The potato cultivars may be different in terms of physical, chemical and functional properties, processing potentials, as well as post-processing parameters [1]. Heat processing in water (boiling) or oil (frying) before the consumption of potatoes causes the gelatinization of the starch resulting in the softening of the inner core [2]. The sensory properties, such as, for example, hardness, adhesiveness, cohesiveness, graininess, mealiness, moistness of boiled potato may differ depending on cultivar. Furthermore, the sensory texture quality of cooked potatoes may be predicted by image analysis of raw potatoes [5].

A machine vision with the grading system may be used for automated inspection of potatoes. Postharvest classification and sorting may be important for ensuring consumer expectations in terms of potato quality. Potato grading is commonly performed based on the shape, size, and defects of tubers [6]. Imaging can be performed using various devices, such as, for example, CCD camera, hyperspectral camera, ultraviolet camera, NMR, X-ray CT. It may allow the evaluation of the external, as well as internal features of potatoes [7, 8]. Computer vision ensures high accuracy with low costs, high repeatability, and flexibility. Due to these facts, inspection systems that use machine vision can be successfully applied in modern manufacturers, for example, as a part of a food processing plant for real-time quality control [9]. Systems using machine vision can be used besides other techniques and methods of the identification of potato cultivar, for example, involving human experts, molecular markers, spectroscopy, some of which may be subjective with a higher risk for human error, time-consuming, more expensive or require sophisticated laboratory facilities [3, 4, 10].

The identification of cultivar may be a significant factor at each step in the agricultural production chain. It can be important, among others, for farmers, growers, variety registration agencies, plant breeders, processors, bulk handlers, marketers, end-users [3]. For example, in industrial usages, for the production of potato chips, identification of the cultivar of potato tubers is very important for further cutting and frying [3]. However, after cutting the potatoes intended for the production of chips, the identification of cultivar may also be necessary because different potato cultivars may require different frying conditions and mixing of slices belonging to different cultivars could affect the final product (potato chips). Therefore, the development of models for the cultivar determination of raw potato slices based on slice features may have practical application to verify cultivar authenticity, detect adulteration and avoid cultivar mixing in potato processing. According to Bărăscu et al. [11], the potato cultivars can be grouped into potato cooking types with different usages, such as A, A/B with properties suitable for salad and other dishes; B, B/C for multiple uses; C for chips, French fries; D for starch and alcohol industries. Thus, cooked potato subjected to further processing also should belong to the appropriate potato cooking type for a specific application. Therefore, identifying the cultivar in an intermediate step of meal preparation can also prove useful.

The aim of this study was to evaluate the effect of potato boiling on the correctness of cultivar discrimination. The discriminative models based on textures of slice images of raw and boiled potatoes were developed. Due to the application of image analysis, the research was performed in an objective, inexpensive and fast manner.

Materials and methods

Materials

The tubers of potatoes of three cultivars: ‘Colomba’, ‘Irga’ and ‘Riviera’ were applied in the experiments. The samples of potatoes with a size larger than 35 mm were collected from north-eastern Poland. The studies were performed with the use of two groups of potatoes: raw and processed. Raw potatoes were washed and cleaned before the experiment. The potatoes intended for processing were also washed and cleaned, and then whole potatoes were subjected to boiling. The potatoes were cooked in boiling water until tender for 25 min. After cooking, the potatoes were air-dried and cooled down at room temperature. The slices of raw and processed potatoes were obtained by cutting each potato with a knife. The slices were cut from the middle part of the tubers. One hundred raw potatoes and 100 boiled potatoes of each cultivar were tested. For each potato, one slice was selected and subjected to tests. The image analysis was carried out with the use of 100 slices (one slice from each potato) of raw and 100 slices (one slice from each potato) of processed potatoes of each cultivar, as follows:

  • 100 slices of raw potato tuber ‘Colomba’,

  • 100 slices of raw potato tuber ‘Irga’,

  • 100 slices of raw potato tuber ‘Riviera’,

  • 100 slices of processed potato tuber ‘Colomba’,

  • 100 slices of processed potato tuber ‘Irga’,

  • 100 slices of processed potato tuber ‘Riviera’.

Image analysis

The image acquisition system consisted of a flatbed scanner and a computer. The imaging device (scanner) was connected to the computer. The slice images of raw and processed potatoes were acquired using the Epson Perfection flatbed scanner. The slices were scanned against a black background at the 800 dpi resolution. The obtained images were saved in TIFF format. The slice images of potatoes were subjected to analysis with the use of the Mazda software (Łódź University of Technology, Institute of Electronics, Poland) [12]. The available version of the Mazda software allowed to convert the images to individual color channels R, G, B, L, a, b, X, Y, Z, U, V and S. The original images and images from B, b and Z color channels of slices of potatoes ‘Colomba’, ‘Irga’ and ‘Riviera’ are presented in Fig. 1 for raw potatoes and in Fig. 2 for processed potatoes. The images from B, b and Z color channels are presented in figures because the differences between the slices of individual cultivars were most visible in the case of images from these channels and the discriminative models built using textures selected from these color channels provided the highest accuracies.

Fig. 1
figure 1

The original images and images from different color channels of slices of raw potatoes

Full size image
Fig. 2
figure 2

The original images and images from different color channels of slices of processed potatoes

Full size image

The regions of interest (ROIs) for images were determined. One ROI included one whole slice of potato (flesh without skin). The black background of the slice images facilitated the determination of ROIs. The ROI defined as a set of pixels of slice was clearly separated from the black background by segmenting the image into brightness regions using the manually determined brightness threshold. Thus, the ROI including a lighter slice than the black background was determined. For each ROI in each color channel about 200 textures of the outer surface of the image based on the co-occurrence matrix, run-length matrix, histogram, gradient map, Haar wavelet transform and autoregressive model were calculated [12].

Statistical analysis

The cultivar discrimination of potatoes was carried out using the WEKA 3.9 application (Machine Learning Group, University of Waikato) [13]. A discriminant analysis was performed separately for raw and boiled potatoes and the results were compared to determine the effect of processing on the discriminative power of the textural features of the outer surface of slice images for distinguishing the potato cultivars. The selection of attributes (textures with the highest discriminative power) was carried out with the use of the Best First with the CFS (correlation-based feature selection algorithm) subset evaluator. The analysis was carried out in several steps. In the first step, the discrimination was performed using models including a set of textures selected from all color channels (R, G, B, L, a, b, X, Y, Z, U, V, S). In the next step, the analysis was carried out for RGB, Lab and XYZ color spaces. In this case, the texture parameters were selected separately for each color space. The following step included building discriminative models for individual channels from each color space (channels R, G, B for RGB color space; channels L, a, b for Lab color space; channels X, Y, Z, for XYZ color space). The results for B, b, and Z color channels were included in this paper because the classification accuracies based on textures selected from these color channels were the highest. Additionally, the discriminative model was built for a set combining the textures selected from three color channels (B, b and Z). The discriminant analysis was carried out using tenfold cross-validation mode and the classifiers from Bayes (Bayes Net), functions (multilayer perceptron) and Lazy (IBk) groups [14].

Results

In the case of raw and boiled potatoes, the accuracies of cultivar discrimination were computed based on confusion matrices. The results were presented in the form of confusion matrices for three potato cultivars containing a number of cases that were correctly and incorrectly classified with an indication of average accuracy (%). Additionally, discrimination accuracies (%) for pairs of potato cultivars: ‘Irga’ vs. ‘Colomba’, ‘Irga’ vs. ‘Riviera’, ‘Colomba’ vs. ‘Riviera’ were determined. The raw potatoes of three cultivars ‘Colomba’, ‘Irga’ and ‘Riviera’ were discriminated with high accuracy reaching 94.33% for the Bayes Net and IBk classifiers based on discriminative models including a set of textures selected from all color channels (R, G, B, L, a, b, X, Y, Z, U, V, S) of slice images (Table 1). In the case of the Bayes Net classifier, all cases of ‘Irga’ were correctly included in the class ‘Irga’. Among the cases of potatoes ‘Colomba’, 92 cases were correctly classified as ‘Colomba’, and eight cases were incorrectly included in the class ‘Riviera’. For potatoes ‘Riviera’, 91 cases were correctly classified as ‘Riviera’ and nine cases—incorrectly as ‘Colomba’. In the case of the IBk classifier, also all cases of ‘Irga’ were correctly classified. For ‘Colomba’, 90 cases were correctly classified as ‘Colomba’, and ten cases as ‘Riviera’. For ‘Riviera’, 93 cases were correctly included in the class ‘Riviera’ and seven cases—incorrectly in the class ‘Colomba’. Considering the accuracy of discrimination performed for pairs of potato cultivars, ‘Irga’ vs. ‘Colomba’, as well as ‘Irga’ vs. ‘Riviera’ were distinguished with 100% correctness. Potatoes ‘Colomba’ vs. ‘Riviera’ were discriminated with the accuracy of up to 91.50% (Bayes Net and IBk classifiers).

Table 1 The confusion matrix and discrimination accuracy of raw potato slices of different cultivars based on a set of textures selected from all color channels (R, G, B, L, a, b, X, Y, Z, U, V, S)
Full size table

In the case of discriminative models built for color spaces, the highest accuracy of discrimination of raw potato ‘Colomba’, ‘Irga’ and ‘Riviera’ was equal to 94% for Lab color space (Bayes Net classifier) (100 cases of ‘Irga’ were correctly classified; 91 cases belonging to ‘Colomba’ were correctly included in the class ‘Colomba’ and nine cases in the class ‘Riviera’; 91 cases of ‘Riviera’ were correctly classified as ‘Riviera’ and nine cases—as ‘Colomba’) and XYZ color space (Multilayer Perceptron classifier) (100 cases belonging to ‘Irga’ were correctly classified; 93 cases of ‘Colomba’ were classified as ‘Colomba’ and seven cases—as ‘Riviera’; 89 cases of ‘Riviera’ were classified as ‘Riviera’ and 11 cases—as ‘Colomba’) (Table 2). In the case of textures selected from RGB color space, the average accuracy for discrimination of three potato cultivars reached 91.33% (multilayer perceptron). All cases of ‘Irga’ were correctly classified. Potatoes ‘Colomba’ were correctly classified in 86% (86 cases from 100) and ‘Riviera’—in 88% (88 cases from 100). For pairs of cultivars, ‘Irga’ vs. ‘Colomba’ and ‘Irga’ vs. ‘Riviera’ were characterized by 100% discrimination accuracy in the case of all color spaces and all classifiers. For ‘Colomba’ vs. ‘Riviera’, the correctness reached 91% (Lab color space, Bayes Net classifier and XYZ color space, Multilayer Perceptron).

Table 2 The confusion matrix and discrimination accuracy of raw potato slices of different cultivars based on textures selected from RGB, Lab and XYZ color spaces
Full size table

For discriminative models built based on textures selected from color channels (Table 3), the accuracies were lower than for color spaces (Table 2). In the case of individual color channels, the highest average discrimination accuracy for three potato cultivars was equal to 92% and was observed for channel b and the Bayes Net classifier, from which all slice images of potatoes ‘Irga’ were correctly classified, 87 cases of ‘Colomba’ were classified as ‘Colomba’ and 13 cases as ‘Riviera’, and 89 cases of ‘Riviera’ were correctly classified and the remaining 11 cases were classified as ‘Colomba’ (Table 3). Both, channels B and Z provided the average correctness of up to 86% for multilayer perceptron. The potatoes ‘Irga’ were completely correctly classified. The lowest accuracies were observed for potatoes ‘Riviera’. In the case of discriminative models built based on textures selected from color channel B, only 78 from 100 cases of ‘Riviera’ were correctly classified and for color channel Z—77 from 100 cases of ‘Riviera’ were included in class ‘Riviera’. Combining the selected textures from channels B, b and Z resulted in an increase in discrimination accuracy to 92.33% for three cultivars and 88.5% for ‘Colomba’ vs. ‘Riviera’. In the case of discrimination of ‘Irga’ vs. ‘Colomba’ and ‘Irga’ vs. ‘Riviera’, the accuracy of 100% was obtained (Table 3).

Table 3 The confusion matrix and discrimination accuracy of raw potato slices of different cultivars based on textures selected from color channels B, b, Z and combined set of textures selected from channels B, b and Z
Full size table

Very high accuracies of discrimination of processed (boiled) potatoes ‘Colomba’, ‘Irga’ and ‘Riviera’ were observed using models built based on a set of textures selected from all color channels (R, G, B, L, a, b, X, Y, Z, U, V, S). The accuracies were equal to 98.67% for the Multilayer Perceptron classifier, 98.33% for Bayes Net and 97.67% for IBk (Table 4). These results are higher than for discriminative models based on textures selected from all color channels built for three cultivars (‘Colomba’, ‘Irga’ and ‘Riviera’) of raw potatoes, for which the correctness reached 94.33% (Table 1). In the case of boiled potato, for the multilayer perceptron and Bayes Net classifiers, 100 cases from the actual class ‘Colomba’ were correctly included in predicted class ‘Colomba’ and 100 cases from class ‘Irga’ were correctly classified as ‘Irga’. A discriminant analysis of pairs of potato cultivars revealed the complete distinction of potatoes ‘Irga’ from ‘Colomba’ and ‘Irga’ from ‘Riviera’ for all classifiers. Potatoes ‘Colomba’ and ‘Riviera’ were discriminated with an accuracy of up to 98% (multilayer perceptron), for which slices of ‘Colomba’ were correctly classified in 100%, and 96 cases belonging to cultivars ‘Riviera’ were included in class ‘Riviera’ and four cases—in class ‘Colomba’.

Table 4 The confusion matrix and discrimination accuracy of boiled potato slices of different cultivars based on a set of textures selected from all color channels (R, G, B, L, a, b, X, Y, Z, U, V, S)
Full size table

Also, high discrimination accuracies for boiled potato of three cultivars were obtained for color spaces (Table 5). The highest accuracy was equal to 98% for RGB color space and the Multilayer Perceptron classifier. Discriminative models built based on textures selected from Lab color space provided an accuracy of up to 97% (Bayes Net), and for models built using textures from XYZ color space, the correctness was equal up to 97.67% (multilayer perceptron). In the case of all color spaces and all classifiers, the potatoes ‘Irga’ were completely distinguished from other potato cultivars (‘Colomba’, ‘Riviera’). The pair of cultivars ‘Colomba’ and ‘Riviera’ was discriminated with the correctness of up to 97% in the case of RGB color space and multilayer perceptron classifier.

Table 5 The confusion matrix and discrimination accuracy of boiled potato slices of different cultivars based on textures selected from RGB, Lab and XYZ color spaces
Full size table

In the case of discriminant analysis performed based on textures selected for individual color channels (Table 6), the highest accuracy of 95.33% was observed for color channel b and Multilayer Perceptron classifier. The slice images of potatoes ‘Irga’ were characterized by the 100% accuracy of discrimination. Potatoes ‘Colomba’ were correctly classified in 95% (95 cases of ‘Colomba’ were classified as ‘Colomba’, and five cases as ‘Riviera’). Potatoes belonging to cultivar ‘Riviera’ were classified with an accuracy of 91% (91 cases of ‘Riviera’ were classified as ‘Riviera’, and nine cases as ‘Colomba’). For comparison of pairs of potato cultivars, the accuracy of 100% was noted between ‘Irga’ and ‘Colomba’, as well as ‘Irga’ and ‘Riviera’, and the accuracy of up to 93% (color channel b, Multilayer Perceptron) was determined for ‘Colomba’ vs. ‘Riviera’. The increase in the correctness of discrimination was achieved by combining the textures selected from channels B, b, and Z (Table 6). The accuracy reached 96.67% for the Bayes Net classifier, from which potatoes ‘Irga’ were correctly classified in 100% cases, and ‘Colomba’, as well as ‘Riviera’—in 95% cases.

Table 6 The confusion matrix and discrimination accuracy of boiled potato slices of different cultivars based on textures selected from color channels B, b, Z and combined set of textures selected from channels B, b and Z
Full size table

Discussion

The obtained results revealed that the boiling resulted in an increase in cultivar discrimination of potato slices. In the case of the discrimination performed using models built for a set of textures selected from all color channels (R, G, B, L, a, b, X, Y, Z, U, V, S), for color spaces and color channels, the accuracies were the higher for the boiled than raw potatoes. Additionally, the usefulness of textures of the outer surface of slice images of potato for cultivar discrimination of potato tuber was proved. Based on literature data, research on the image analysis of potatoes focused mainly on whole tubers. Machine vision combined with artificial neural networks was used by Azizi et al. [3] for the discrimination of ten potato cultivars based on color, morphological and textural features of whole tubers. The obtained accuracy ranged from 40 to 93.3% depending on the cultivar for stepwise discriminant analysis (DA). Using neural network analysis, the correctness increased up to 100%. Mercurio and Hernandez [15] applied image processing combined with a convolutional neural network with the use of size and skin properties to classify sweet potatoes belonging to five cultivars achieving an average accuracy of 96.33% (from 95 to 98.33% for individual cultivars). Cultivar identification of sweet potato using NIR hyperspectral imaging and FT-MIR microspectroscopy revealed that two cultivars may be distinguished with 100% correctness [16]. Identification of two cultivars of potato tubers using neural image analysis based on aspect factors, geometric and color features was successfully performed with the quality of testing of above 0.99 by Przybył et al. [17]. In addition to cultivar identification, image analysis based on morphological features was also applied for the detection of potatoes with an irregular shape [18], the detection of potato tubers with damages, diseases and defects [19,20,21,22], the prediction of shape parameters and mass of normal and deformed potatoes [7], the classification of the potato tubers based on moisture content [23], the prediction of changes in moisture content and color of sweet potato during drying [24], the prediction of texture and color in cooked and cooked freeze-dried rehydrated potatoes [25]. The own studies regarding the evaluation of the effect of boiling on cultivar discrimination of potato based on the textural features of the outer surface of slice images completed the literature data on image analysis of potato. The results demonstrated the usefulness of slices in addition to whole tubers for potato cultivar discrimination. Further research may include the development of models based on texture parameters of slices for cultivar discrimination of potatoes subjected to other methods of processing. Further studies can also be intended to evaluate the robustness of classifiers by examining samples of the same potato cultivars collected from different regions and different growing seasons.

Conclusions

The obtained results revealed that the analysis of slice images acquired with the use of a flatbed scanner allows for cultivar discrimination of raw and processed potatoes in an objective, inexpensive and fast manner with high accuracy. The discriminative models built based on a set of textures selected from color channels R, G, B, L, a, b, X, Y, Z, U, V, S produced the highest results, 94.33% for raw potatoes ‘Colomba’, ‘Irga’ and ‘Riviera’, and 98.67% for processed potatoes. These models revealed that both in the case of raw and processed potatoes, 100% discrimination accuracy was achieved between potatoes ‘Irga’ vs. ‘Colomba’ and ‘Irga’ vs. ‘Riviera’. Raw potatoes ‘Colomba’ vs. ‘Riviera’ were distinguished with the correctness of up to 91.50% and processed potatoes ‘Colomba’ vs. ‘Riviera’—of up to 98%. The results may be applied in practice to verify the cultivar authenticity of sliced potato, detect adulteration, and avoid cultivar mixing in potato processing. The identification of cultivar may be a significant factor at each step in the agricultural production chain. It can be important, among others, for farmers, growers, cultivar registration agencies, plant breeders, processors, bulk handlers, marketers, end-users. Further research can be intended to develop the discriminative models for potatoes subjected to other methods of processing and can relate to examining potato samples collected from different growing seasons and several regions.