Home The application of convolutional neural networks, LF-NMR, and texture for microparticle analysis in assessing the quality of fruit powders: Case study – blackcurrant powders
Article Open Access

The application of convolutional neural networks, LF-NMR, and texture for microparticle analysis in assessing the quality of fruit powders: Case study – blackcurrant powders

  • Krzysztof Przybył EMAIL logo , Katarzyna Samborska , Aleksandra Jedlińska , Krzysztof Koszela , Hanna Maria Baranowska , Łukasz Masewicz and Przemysław Łukasz Kowalczewski
Published/Copyright: August 22, 2025
Download Article (PDF)
Become an author with De Gruyter Brill
Submit Manuscript Author Information
REVIEWS ON ADVANCED MATERIALS SCIENCE
From the journal REVIEWS ON ADVANCED MATERIALS SCIENCE Volume 64 Issue 1

Abstract

It can be observed that dynamic developments in artificial intelligence contributing to the evolution of existing techniques used in food research. Currently, innovative methods are being sought to support unit processes such as food drying, while at the same time monitoring quality and extending their shelf life. The development of innovative technology using convolutional neural networks (CNNs) to assess the quality of fruit powders seems highly desirable. This will translate into obtaining homogeneous batches of powders based on the specific morphological structure of the obtained microparticles. The research aims to apply convolutional networks to assess the quality, consistency, and homogeneity of blackcurrant powders supported by comparative physical methods of low-field nuclear magnetic resonance (LF-NMR) and texture analysis. The results show that maltodextrin, inulin, whey milk proteins, microcrystalline cellulose, and gum arabic are effective carriers when identifying morphological structure using CNNs. The use of CNNs, texture analysis, and the effect of LF-NMR relaxation time together with statistical elaboration shows that maltodextrin as well as milk whey proteins in combination with inulin achieve the most favorable results. The best results were obtained for a sample containing 50% maltodextrin and 50% maltodextrin (MD50-MD70). The CNN model for this combination had the lowest mean squared error in the test set at 2.5741 × 10−4, confirming its high performance in the classification of blackcurrant powder microstructures.

Graphical abstract

Keywords: convolutional neural networks; low-field nuclear magnetic resonance; texture analysis; scanning electron microscopy; fruit powders; blackcurrant

1 Introduction

Following the development of technology, advanced techniques have been introduced to ensure sustainability, safety, and quality in the production and distribution of food [1]. Food preservation is achieved through traditional processes such as drying, conserving, pasteurization, or freezing [2–4]. Artificial intelligence is becoming a key aspect in the optimization of these processes [5,6]. Different methods of machine learning, including deep learning, will primarily allow to minimize the risk of inefficiently obtained food products, which translates into the reduction of waste of the final product obtained. In the age of Industry 4.0 and 5.0 in food production and processing, to use of advanced methods like image analysis, automation, robotics, artificial intelligence, and data analysis is becoming more effective and improves the quality condition as well as the reproducibility of the obtained food products [4,7–10]. Advanced machine and deep learning techniques allow more efficient management of products, water, and energy in the process of obtaining and processing the above [11]. In addition, the use of machine and deep learning algorithms allows to accurately monitor and control the quality condition of food products at various stages of food processing [12,13].

Nowadays, companies try to offer the market as natural a colorant as possible. As a result of low-temperature spray drying, blackcurrant powders [14–16] were obtained in the research. Fruit powders add flavor and functionality to various products such as teas, jellies, puddings, instant soups, instant drinks, all dairy products, dietary supplements, confectionery, pharmaceutical preparations, etc. In fact, proper preservation of blackcurrant powder will maintain the content of bioactive compounds in the resulting food products and preserve their quality coefficient. It is worth knowing that exposure to high temperatures causes changes in the nutritional value of powders. Deterioration of the quality of products as a result of their storage at room temperature causes degradation of compounds, including the oxidation of vitamins. Effective yet low-energy methods are being sought to prevent the degradation of compounds.

Recent results show that there are few works using deep learning solutions in the context of spray drying [12]. This is due to the fact that spray drying is a complicated process for obtaining fruit powders. The process of spray drying carries the risk of not obtaining homogeneous powders, giving the correct morphological structure to the powder microparticles, and thus obtaining the right quality of the final product. For this reason, it is very important, in addition to the established drying parameters (inlet and outlet air temperatures, determination of air velocity flux and humidity), to adjust the way of controlling the morphological structure of powder microparticles in order to minimize the loss of their quality [17–19]. In another recent research, an attempt was made to use traditional artificial neural network (ANN) methods and convolutional neural networks (CNNs), to identify raspberry powders that differ in the amount and type of polysaccharides [20]. The aim of this work was to develop optimal multi-layer perceptron network and CNNs with optimal classification performance. The new methods used are the result of technological advances, acquired knowledge, as well as experience. The new developments use advanced algorithms, neural network models, and data processing techniques that contribute to better performance and capabilities in analysis, prediction, and decision-making problems in the field of machine and deep learning.

This research aims to fill a gap in the analysis of food quality and preservation by evaluating the morphological structure of fruit powder microparticles using CNNs, which are among the most advanced methods of artificial intelligence, specialized low-field nuclear magnetic resonance (LF-NMR) and instrumental texture analysis [21–24]. The morphology of microparticles can be captured using digital images, allowing for faster analysis of large amount of image data. CNN will effectively analyze individual areas of powder microparticles contained in the image, which will serve as input data, and generate new tensors from them. The application of deep learning based on CNN will determine an effective algorithm for image classification due to the presence of repeated, recognizable, and similar patterns in the image [25]. The image pattern will determine the state of morphological structure for each variant of blackcurrant powder. When determining the pattern, a description characterizing the shape of the microparticle found in the image, the height of the image, the width of the image, and the depth as the number of dimensions of the color space [26] will be used as the input tensor of the network. As part of the comparative analysis of the structure of blackcurrant powders, the use of a specialized LF-NMR method allows us to effectively understand the research samples obtained by analyzing the molecular structure of blackcurrant powders. In addition, the instrumental texture was used in the study to provide a link in the effective identification of blackcurrant powders. LF-NMR and texture analyses combined with CNN provide complete information on the quality, properties, and structure of the analyzed powders.

2 Materials and methods

2.1 Preparation of a solution from blackcurrant concentrate, taking into account the amount and content of the selected carrier, and performing low-temperature spray drying

As part of the implementation of the scientific activity, concentrated blackcurrant juice (blackcurrant concentrate) with a density of 67°Brix and a color coefficient of 2.352 (E520/E420) was purchased in the amount of 5 kg obtained from a batch of 250 kg of this product from the company Białuty Sp. z o.o and carriers such as maltodextrin (MD; Hortimex), gum arabic (GA; Hortimex), inulin (IN; Hortimex), milk whey protein (WPC; Mlekovita), fiber (F; Hortimex), and microcrystalline cellulose (C; Hortimex). Preparation of solutions was done by mixing blackcurrant concentrate with the appropriate amount of water and carriers. Twelve solution variants were prepared (Table 1), which differed in the type and proportion of carrier. The proportion of carriers in the solution was 30 and 50% solids, respectively. The solutions were prepared by mixing the concentrate with the appropriate amount of water and carrier, and then dried in batches of 600 g.

Table 1

Names of the blackcurrant powders were determined by their dependence on the type and proportion of the carrier

Type of carrier The ratio of blackcurrant concentrate to carrier solids
50:50 70:30
Arabic gum (GA) GA50 GA70
Maltodextrin (MD) MD50 MD70
Inulin (IN) IN50 IN70
Whey protein (W) W50 W70
Fiber (F) F50 F70
Microcrystalline cellulose (C) C50 C70

The prepared solutions were dried in a laboratory dryer at an inlet air temperature of 80°C and outlet air temperature of 50°C. During the drying process, the humidity of the air was 0.5 g·cm−3. The feed flow rate was 0.3 mL·s−1. As a result of low-temperature drying, blackcurrant powders were obtained.

2.2 Physical analysis of blackcurrant powders (LF-NMR and instrumental texture)

The analysis was carried out in order to characterize the properties of blackcurrant powders, which allows for effective assessment of their quality. Instrumental texture measurement of blackcurrant powders was carried out using a TA.XTplus Texture Analyzer (Stable Micro Systems, UK) with an SMS P36R attachment. A calibration of the attachment height of 110 mm was performed, 40% compression for samples up to 25 mL was taken into account. The sample weight was 58 ± 0.1 g. For each test variant, eight replicates were performed. A total of 96 teaching cases were obtained in the numerical dataset.

As an additional research, a comparative measurement was performed by measuring the blackcurrant powder variants by LF-NMR. Samples of the analyzed powders were placed in 7 mm diameter measuring tubes. The tubes were sealed with Parafilm®. The volume of the samples was 770 mm3.

The spin–lattice T 1 relaxation times were measured using pulse NMR spectrometer PS15T operating at 15 MHz (Ellab, Poznań, Poland). The inversion-recovery (180°–TI–90°–TR) pulse sequence was applied [27,28]. The 90° radio pulse (RF) pulse was from 2.2 to 2.6 ms. The distance inversion time (TI) between RF pulses was changed from 0.2 to 100 ms, the repetition time repetition time was 15 s. Each time, 32 free induction decay (FID) signals and from 38 to 54 points from each FID signal were collected. Five accumulations of NMR signals were applied each time.

Calculations of the T 1 relaxation time values were performed with CracSpin software (Jagiellonian University, Cracow, Poland) [29] using a “spin grouping” approach. The accuracy of the relaxation parameters was determined, and standard deviations were calculated. The recovery of the magnetization was described by the equation given below [29]:

(1) M z ( TI ) = M 0 1 − 2 e − TI T 1 ,

where M z (TI) is the actual magnetization value and M 0 is the equilibrium magnetization value.

Monoexponential magnetization recovery, observed in this study, indicated that the system relaxed within one TI spin–lattice relaxation time. The measurements were performed in controlled temperature at 20.0 ± 0.5°C. As a result, 12 cases of each research variant were acquired.

2.3 Microscope image acquisition and processing

The next step required preparing blackcurrant powder samples for morphological structure determination using a scanning electron microscope (SEM). The samples were attached to the sample section with double-sided adhesive tape and cathodically sputtered with gold. The instrument was calibrated with a secondary electron detector before testing. The working distance of the detector from the samples was 10 mm inside (WD = 10 mm). The accelerating voltage for each sample was 12.82 kV. There were 35 repetitions in the form of microscope photographs for each variant of blackcurrant powders. Photographs of microparticles were obtained at 500 times magnification at a scale of 100 μm (a total of 420 microscope images).

After the problem was formulated in machine learning, preprocessing of microscope images with a resolution of 2,048 × 1,576 (the original images) and their format with a.TIFF extension was carried out. For this purpose, the free Python Imaging Library (PIL) was used [30,31]. The PIL library is one of the most popular tools for image processing using the Python language. The PIL library allows for automated yet fast image conversion, point transformations, batch processing, filtering, bit depth transformation, and changing the color space of images. In technical terms, the goal of image processing is to skillfully acquire a pattern, adjust filtering, transformations, storage format changes, and image resolution for better quality of the resulting image. In view of the above, image processing seems to be important in the field of machine learning in the aspect of computer vision to improve the quality and extract the necessary information from the established pattern of images.

In the first step of the study, primary images with a.TIFF extension were converted to secondary images with a.jpg extension. The converted images were saved under a new name. The next step was to do segmentation of the images. The number of segments along the x-axis and y-axis was calculated, which made it possible to perform cropping of the secondary images by the indicated step relative to the x- and y-axis parameters. In order to automate the process, the batch cropping of these images was performed again using the PIL library. The result was 1,400 × 1,400 resolution secondary images, the aim of which was to obtain a consistent view of the data. The final step was to perform a point transformation on the images. The segmented secondary images were subjected to 90°, 180°, and 270° transformations and stored in the output catalog. As a result of data pre-processing, 105 digital images corresponding to the learning set were obtained for each research variant of blackcurrant powders. The purpose of pre-processing of secondary images obtained by image processing, in addition to the qualitative measures mentioned above, becomes the normalization of the input data in the learning set. The rescaling of the pixel values of the images in the image analysis will improve the quality of learning the designed neural models in the machine learning process.

2.4 Design and learning of convolutional networks

This step required the formulation of the variables in the sets and the selection of a machine-learning algorithm to design the model and carry out the learning process to obtain an adequate determination index of blackcurrant powders. Two learning and test sets were prepared for each research variant. An important step in preparing the sets is to determine the split between the learning set and the test set. In computer vision machine learning, the split is assumed to be 60:40, 70:30, or 80:20 [32,33]. In the work averaging, the division of the sets was made at a ratio of 75:25. The learning set contained 105 images processed through data pre-processing. The test set contained 35 microscope images with a fixed resolution of 1,400 × 1,400, which were not in the learning set. In the test samples, using microscope images (SEM), 36 learning sets were finally created, which represented combinations with the lowest (30%) (total n = 15) and highest (50%) (total n = 15) proportion of the selected type (total n = 6) of carrier in the blackcurrant solution.

The design of convolutional networks was performed using 36 variants of blackcurrant powder sets. The goal of this stage was to develop CNNs. In the design of CNNs using the Tensorflow and KERAS libraries, the sequential model is used, which is one of the simplest and most popular ways of building models in the library [34]. The sequential model is a linear application of layers, meaning that each layer has one input and exactly one output, and data flow between convolutional layers, activation function, and pooling layers [35]. The structure of the designed convolutional networks for each learning set included: convolution layers (Convolution) to determine the size of the input tensor, max pooling layers (Max-Pooling) to reduce the size of the input tensor, flattening layer (Flatten) to transform the image to a vector, and dense layers as densely connected. In research the sequential CNN model consisted of eight layers: convolutional layers (Conv2D) and maximum pooling layers (MaxPooling2D). The CNN model also included a flattening layer to transform the multidimensional input data (image pixels of blackcurrant powder) into a one-dimensional vector. The model also used a dense layer of 288 neurons, meaning that each neuron was connected to the input of the previous layer. This layer used the ReLU activation function, which introduces non-linearity into the network, allowing the model to better represent the complexity of the data. The output layer consisted of a neuron and a sigmoidal activation function, which transformed the result into a value between 0 and 1, accurately representing the blackcurrant varieties. The structure is shown in Figure 1. The initialization of the CNN used a loss function of the “binary_crossentropy” type, applied to this research as binary classification problem. The “adam” learning algorithm was used, which is the most popular and probably the most efficient optimization algorithm for updating weights during CNN learning [36–38].

Figure 1 CNN model structure in blackcurrant powders.
Figure 1

CNN model structure in blackcurrant powders.

An essential step in the design of suitable neural models for microscopic images identification was the analysis between overfitting and underfitting. In order to avoid over-fitting of the CNN model to the learning set and also to effectively use these networks in image recognition, a scaling of the data was performed [39]. This involved the creation of additional convolutional layers for the input tensor, resulting in a balance between the fit and generalization ability of the CNN models.

2.5 Statistical analysis

The research article presents results using analysis of variance (ANOVA) [40,41] and principal component analysis () [42–44] methods. A Tukey test was applied for each sample with a significance level of p = 0.05 [45,46]. Python and Statistica 13.3 were used to perform statistical analysis. Using the Tukey test after ANOVA, differences between blackcurrant powder variants were statistically determined in Python. The scipy and scikit-learn libraries were used within the Python environment [47,48]. For the PCA method, relationships between textural variables were identified in Statistica 13.3 (TIBCO Software Inc., Palo Alto, CA, USA).

3 Results

3.1 Preparation of a solution from blackcurrant concentrate, taking into account the amount and content of the selected carrier, and performing low-temperature spray drying

As a result of dehumidified air-assisted drying, blackcurrant powders were obtained according to the procedure of Przybyl et al. [15]. Table 1 shows different combinations of carriers, i.e., gum arabic, maltodextrin, inulin, whey milk proteins, fiber, microcrystalline cellulose, along with their specific proportion in solution, i.e., the smallest (30% content) and the largest (50% content). As a result, 12 variants of blackcurrant powders were obtained.

3.2 Physical analysis of blackcurrant powders (LF-NMR and instrumental texture)

According to statistical analyses using the Tukey test, PCA, and correlation methods, the relationships between textural parameters and T 1 relaxation time were compared using the LF-NMR method. T 1 relaxation time and textural coefficients in blackcurrant powders made it possible to assess their quality, porosity, homogeneity, and the degree of structure in which individual microparticles are bound together. The correlation between the textural variables is intended to show the degree of interrelatedness (Figure 2). Table 2 shows the textural coefficients and T 1 relaxation time from the LF-NMR method. It turned out that maltodextrin (50% share) and milk whey proteins (30% carrier content) were the most statistically significantly different from the other test variants of blackcurrant powders. The average values shown in Table 2 for each homogeneous group showed that maltodextrin and whey proteins are significantly different from the remaining coefficients.

Figure 2 PCA (a) combined with correlation (b) of textural variables.
Figure 2

PCA (a) combined with correlation (b) of textural variables.

Table 2

Texture coefficients and spin–lattice relaxation times of blackcurrant powders containing 50% and 30% carrier gum arabic (GA), inulin (IN), maltodextrin (MD), whey protein (W), fiber (F), and microcrystalline cellulose (C) by dehumidified air-assisted spray drying

T 1 Force Hardness Adhesiveness Springiness
MD50 70.66 ± 2.46 4547.55 ± 538.24a 8885.17 ± 1356.42ab −9.87 ± 9.93b 0.57 ± 0.24ba
MD70 68.20 ± 3.58 5435.31 ± 1614.63ab 11007.98 ± 3328.79ab −3.63 ± 3.64b 0.49 ± 0.31ba
GA50 115.41 ± 10.80 18160.68 ± 3804.28d 32912.67 ± 6195.51d −9.28 ± 4.81b 0.74 ± 0.22b
GA70 71.95 ± 5.79 10733.51 ± 4385.58c 20781.50 ± 7769.08c −4.37 ± 2.39b 0.47 ± 0.09ba
IN50 93.43 ± 10.55 5628.67 ± 1480.43ab 10932.30 ± 3003.33ab −5.65 ± 7.16b 0.45 ± 0.35ba
IN70 58.31 ± 6.26 6387.79 ± 2638.03ab 11788.26 ± 4491.27ab −1.87 ± 1.49b 0.35 ± 0.15a
C50 53.32 ± 4.50 5672.24 ± 1161.70ab 11167.43 ± 2188.69ab −5.53 ± 1.12b 0.36 ± 0.15a
C70 93.29 ± 3.30 5997.64 ± 1814.92ab 12244.32 ± 3348.27ab −9.75 ± 3.66b 0.78 ± 0.33b
F50 44.76 ± 5.26 9401.71 ± 3930.57bc 16668.08 ± 8541.36bc −26.90 ± 10.57a 0.82 ± 0.21b
F70 86.69 ± 2.90 6378.40 ± 1288.15ab 11734.90 ± 2612.61ab −3.65 ± 2.66b 0.49 ± 0.20ba
W50 74.97 ± 8.30 5890.29 ± 2272.78ab 8742.09 ± 4249.09a −23.40 ± 10.14a 0.47 ± 0.17ba
W70 71.19 ± 5.30 4615.75 ± 1277.42a 6603.19 ± 3164.72a −23.06 ± 7.20a 0.50 ± 0.13ba
Cohesiveness Gumminess Chewiness Resilience
MD50 0.13 ± 0.02a 1181.10 ± 186.94a 680.13 ± 334.10a 0.12 ± 0.01ba
MD70 0.14 ± 0.01ab 1579.05 ± 503.14a 747.83 ± 527.33a 0.14 ± 0.01bc
GA50 0.21 ± 0.01abc 6816.16 ± 1201.99c 5161.88 ± 2049.06c 0.23 ± 0.02g
GA70 0.15 ± 0.01ab 3127.73 ± 1019.03b 1444.85 ± 552.23a 0.15 ± 0.02cd
IN50 0.13 ± 0.01a 1387.29 ± 361.55a 657.45 ± 614.32a 0.12 ± 0.01ba
IN70 0.15 ± 0.02ab 1796.63 ± 828.21a 596.40 ± 270.52a 0.13 ± 0.01bac
C50 0.14 ± 0.01ab 1581.53 ± 289.09a 555.18 ± 180.14a 0.13 ± 0.01bac
C70 0.14 ± 0.01ab 1757.94 ± 621.35a 1470.84 ± 874.29ab 0.14 ± 0.01bc
F50 0.23 ± 0.04bcd 3471.43 ± 958.24b 2778.30 ± 806.74b 0.18 ± 0.03de
F70 0.19 ± 0.02ab 2268.24 ± 542.67ab 1055.32 ± 314.18a 0.19 ± 0.02e
W50 0.29 ± 0.14cd 2356.36 ± 986.97ab 1242.89 ± 650.15a 0.12 ± 0.02ba
W70 0.30 ± 0.10d 1734.71 ± 397.17a 868.41 ± 337.69a 0.10 ± 0.01a

a–g: the differences between mean values with the same letter in columns without column T 1 were statistically significant (p < 0.05). Column T 1 corresponds to the resulting values using the LF-NMR method.

In the study of textual analysis, cohesiveness, springiness, chewiness, gumminess, resilience, force (Force_1), hardness, and adhesiveness were determined for blackcurrant powders. Maltodextrin showed the lowest Force_1 at 4547.55, and this translates into regularity, linearity in the microparticle structure of this blackcurrant powder variant. As Force_1 increases, their level of texture in blackcurrant powders changes, causing irregularities in the structure or increased inter-particle spaces. In the case of the hardness coefficient, for which the qualitative analysis of the powders affects their degree of granularity [49] was also statistically shown that the maltodextrin carrier was one of the three to achieve the lowest value at 8885.17 relative to the other test variants. The lowest hardness of the powders was shown for the milk whey protein carrier (6603.19) and the highest hardness was achieved by the blackcurrant powder variant with gum arabic (32912.67). Blackcurrant powders with a higher hardness coefficient may tend to form the so-called clumps in the morphological structure of microparticles [49].

Moreover, the higher the hardness of coefficient, the more favorable the flow, spreadability, and stability of powders. Adhesiveness, on the other hand, has a significant effect on the solubility of powders. The higher the adhesiveness index, the more difficult it is for the carrier to dissolve in solutions. The results in Table 2 relating to adhesiveness statistically show that whey protein and microcrystalline cellulose have the lowest solubility index in food products. In the case of springiness or resilience, which determines the ability of powders to return to their original shape, the different research variants do not differ statistically in their values. Nevertheless, their low content indicates lower elasticity, which means that they do not return to their original shape. Cohesiveness as the primary determinant of intermolecular bonding of powders from the point of view of microparticle structure, has an impact on their cohesion [50,51]. Within the individual variants of powders, the cohesion index was determined to be significantly low. This means that blackcurrant powders have a lower tendency to form agglomerates and lumps [52]. Low cohesiveness significantly affects scattering and the microparticles will allow them to separate more efficiently.

Proper storage of these powders is essential. In the case of the coefficient of gumminess, which is responsible for the elasticity of the powders, it was shown that the highest elasticity, viscosity, was achieved by the powder variant with arabic gum (GA50) with a value of 6816.16 and the lowest with maltodextrin (MD50) with a value of 1181.10. Chewiness in instrumental texture refers to the degree of difficulty in breaking and chewing the food product. Very low chewiness was again achieved by maltodextrin and the highest value by arabic gum.

As part of the comparative analysis between the instrumental texture results, the relationships of relaxation times were determined [24,53], which illustrated the molecular movements of water in blackcurrant powders. The LF-NMR method made it possible to find a specific range of temperatures at which the formation of lattice nodes occurs for the different variants of blackcurrant powders. The longest T 1 relaxation time by LF-NMR was achieved by the variant with gum arabic carrier (GA50), while the shortest T 1 was achieved by the variant with fiber (F50). In practical terms, this means that the longer the T 1 relaxation time obtained with LF-NMR, the more ordered the structure of the blackcurrant powder microparticles becomes (a more solid-body structure).

The correlation and PCA result (Figure 2a and b) when evaluating the dimensionality reduction of the data and attempting to extract significant data, it can be found that cohesiveness is strongly negatively correlated with adhesiveness giving a correlation coefficient of R 2 = −0.7. This means that the strong negative correlation between cohesiveness and adhesiveness determines the greatest variability between each other. The principal component of PC1 reflects the direction in which there is the greatest correlation between cohesiveness and adhesiveness. In the case of PC2, this measure shows the opposite direction of the relationship between these data. It was observed that Force_1 shows a strong correlation between hardness R 2 = 0.98 and resilience R 2 = 0.96. This is due to the fact that the strong correlation between these variables affects the variability of the principal components. The variable Force_1, which determines the effect of the texture on the morphological structure of blackcurrant powder microparticles, responds to the other variables hardness and resilience by determining, among other things, the variability of microparticle shape. Considering the aforementioned variable hardness shows a strong correlation with the variable gumminess (R 2 = 0.91), and the variable gumminess shows a strong correlation with chewiness (R 2 = 0.88). This expresses common information about the powder microparticle that significantly influences the principal components in the PCA.

3.3 Microscope image acquisition and processing

Microscopic data collection using SEM resulted in 12 sets of images describing blackcurrant powders (Figure 3a–l). Each variant stored information about the morphological structure of the microparticles (Figure 3).

Figure 3 Microparticles from 12 variants of blackcurrant powders containing 50%,and 30% carrier microcrystalline cellulose (C) (a) and (b), maltodextrin (MD) (c) and (d), whey protein (W) (e) and (f), fiber (F) (g) and (h), gum arabic (GA) (i) and (j), and inulin (IN) (k) and (l) which were isolated from microscope image (SEM).
Figure 3 Microparticles from 12 variants of blackcurrant powders containing 50%,and 30% carrier microcrystalline cellulose (C) (a) and (b), maltodextrin (MD) (c) and (d), whey protein (W) (e) and (f), fiber (F) (g) and (h), gum arabic (GA) (i) and (j), and inulin (IN) (k) and (l) which were isolated from microscope image (SEM).
Figure 3

Microparticles from 12 variants of blackcurrant powders containing 50%,and 30% carrier microcrystalline cellulose (C) (a) and (b), maltodextrin (MD) (c) and (d), whey protein (W) (e) and (f), fiber (F) (g) and (h), gum arabic (GA) (i) and (j), and inulin (IN) (k) and (l) which were isolated from microscope image (SEM).

3.4 Design and learning of convolutional networks

Under the designed CNN models and taking into account the criterion set in the methodology, i.e., over-fitting by under-fitting analysis, and a fixed number of epochs equal to 100 during the learning of each CNN model, it was possible to obtain the results that are shown in Table 3. In Figure 4, the process of obtaining a CNN model is shown, according to the established standard for the design of ANN models in machine learning. Between the variants of blackcurrant powders for which the input data consisted of SEM images, i.e., objects in the form of powder microparticles, the best identified were those with the highest proportion of carriers (50% content) for maltodextrin and inulin (MD50-IN50) (Figure 5a–c). The MD50-IN50 CNN model had the lowest mean square error (MSE) for the test set, which was 8.4583 × 10−4 (Figure 5b), respectively. In comparison, the weakest fit to the data of 50% content and type of carrier in blackcurrant powders was determined by the CNN GA50-W50 model, for which MSE = 2291.7201 × 10−4 (Table 3). Considering the results obtained, those models not included in Table 1 did not meet the criterion of over-fitting and under-fitting relationships, for which the value of the MSE coefficient determined a value above 0.22 and the accuracy coefficient below the fit value at 0.71.

Table 3

Results of blackcurrant powders using CNN

Model Precision MSE (10–4) MAE (10–4) Val_accuracy Val_mse (10–4) Val_mae (10–4) Val_precision Recall Accuracy Val_recall
1 MD50-IN50 1.0000 0.0009 1.2020 1.0000 8.4583 39.0524 1.0000 1.0000 1.0000 1.0000
2 MD50-MD70 1.0000 0.0005 5,417 1.0000 2.5741 45.9237 1.0000 1.0000 1.0000 1.0000
3 GA50-GA70 0.9750 74.7065 184.1010 1.0000 16.1935 117.9243 0.9535 1.0000 1.0000 1.0000
4 W70-IN70 1.0000 0.6201 19 1.0000 0.217 5.4210 1.0000 1.0000 1.0000 1.0000
5 C70-W70 0.9778 73 186 1.0000 1.0196 16 1.0000 1.0000 0.9865 1.0000
6 F50-F70 1.0000 114 316 0.9688 193 360 0.9697 0.9394 0.9750 0.9697
7 F50-MD50 0.9875 130.6599 313.9229 0.9688 288.8454 434.4893 1.0000 0.9750 0.9697 0.9697
8 GA50-IN50 1.0000 14.3871 77.7426 0.9688 159.2872 343.5659 1.0000 1.0000 0.9429 1.0000
9 F50-G50 1.0000 0.0086 2.3117 0.9531 362.9437 517.1319 1.0000 1.0000 1.0000 0.9063
10 C50-GA50 1.0000 0.5910 25.0927 0.9531 338.3946 626.7399 1.0000 1.0000 0.9118 1.0000
11 IN50-IN70 0.9750 92.9790 334.8764 0.9531 431.1070 583.4963 0.9722 0.9722 1.0000 0.9063
12 F70-IN70 0.9767 177 452 0.9531 333 464.0000 0.9394 1.0000 0.9875 0.9688
13 F70-MD70 1.0000 23 92 0.9531 468 620.0000 0.9394 1.0000 1.0000 0.9688
14 MD50-W50 0.9875 67.1487 158.4006 0.9219 686.1994 1162.6472 0.9750 1.0000 1.0000 0.8333
15 MD50-GA50 1.0000 0.0523 4.0701 0.9063 593.6483 908.9045 1.0000 1.0000 1.0000 0.8125
16 W50-IN50 0.9865 85.0857 234.1973 0.8750 944.5102 1343.1749 0.9744 1.0000 0.7949 1.0000
17 W50-W70 1.0000 11.3038 149.9944 0.8438 1138.6122 1738.1969 1.0000 1.0000 1.0000 0.7059
18 C70-IN70 1.0000 33 189 0.8125 1480 2038 0.7209 1.0000 1.0000 1.0000
19 C50-MD50 0.9375 483.5656 806.4995 0.7813 1604.3843 2257.4732 1.0000 0.8611 0.7143 1.0000
20 F70-GA70 1.0000 155 411 0.7812 1716 2020 0.6818 0.9474 0.9750 1.0000
21 GA50-W50 1.0000 0.3772 12.8679 0.7344 2291.7201 2503.0488 1.0000 1.0000 0.6531 1.0000
Figure 4 CNN procedure in blackcurrant powders: SEM image preparation, image preprocessing, creating CNN, tuning model CNN in training, and generate adequate CNN model.
Figure 4

CNN procedure in blackcurrant powders: SEM image preparation, image preprocessing, creating CNN, tuning model CNN in training, and generate adequate CNN model.

Figure 5 Learning charts (accuracy factor), MSE, and MAE for the five highest obtained results from the 21 respective CNNs: MD50-IN50 (a)–(c), MD50-MD70 (d)–(f), GA50-GA70 (g)–(i), W70-IN70 (j)–(l), C70-W70 (ł), (m), and (n).
Figure 5

Learning charts (accuracy factor), MSE, and MAE for the five highest obtained results from the 21 respective CNNs: MD50-IN50 (a)–(c), MD50-MD70 (d)–(f), GA50-GA70 (g)–(i), W70-IN70 (j)–(l), C70-W70 (ł), (m), and (n).

3.4.1 Results of blackcurrant powders using CNN

In the case of results determining the proportion of carrier in the blackcurrant powder carrier variant, the best data recognition rate using the CNN model was achieved with the proportion of maltodextrin or gum arabic carrier. The CNN model (MD50-MD70) (Figure 5d–f) with the lowest proportion of maltodextrin carrier achieved a MSE value of 2.5741 × 10−4 (Figure 5e). In the case of the CNN model (GA50-GA70) (Figure 5g–i), for which the carrier was gum arabic the MSE for the test set was 16.1935 × 10−4 (Figure 5h). Considering the weakest result but at the same time meeting the criterion of CNN model fit was achieved with the CNN model (W50-W70) with the content of milk whey protein carrier.

Considering the type and the smallest proportion of the carrier (30% content) in blackcurrant powder, it turns out that the best CNN learning results were obtained with the participation of milk whey protein carriers with inulin and microcrystalline cellulose (W70-IN70, C70-W70), where the CNN models W70-IN70 (Figure 5j–l) and CNN C70-W70 (Figure 5m, n) achieved MSE values of, respectively, 0.0217 × 10−4 and 1.0196 × 10−4. In comparison, the weakest result for identifying the 30% carrier share between powder variants was achieved by the CNN F70-GA70 model, where the MSE was 1,716 × 10−4.

In Figure 4, it was illustrated the learning results for CNNs characterized by high accuracy, the lowest MSE, and mean absolute error (MAE) in the test set (val_mse, val_mae). The MSE and MAE values suggest that the CNN model predicts the output values well and minimizes discrepancies with the actual labels. Comparing the error functions for the training and test sets, it can be seen that the CNN model shows no clear signs of over-fitting, as confirmed by the stable difference between the errors in the two sets.

These results suggest that the CNN can be effectively used to identify fruit powders, and its accuracy and low error values make it a potentially useful tool in practical applications.

4 Discussion

As a result of the research using microscopic images (SEM), 21 out of 36 CNN models were selected (meeting the over-fitting by under-fitting criterion), which represented combinations with the smallest (30%) and largest (50%) proportion of the selected carrier type in the blackcurrant solution. Based on the results, it can be concluded that it is possible to identify blackcurrant powder microparticles using CNNs. Previous research efforts confirm the effectiveness of artificial intelligence in fruit powder identification [54]. Other studies have investigated the possibility of reducing the carrier content in the production of fruit powders using a spray drying method, with the aim of obtaining fruit powders with a lower carrier content [15]. The current research compared the highest and lowest carrier contents in fruit powders. The aim of the endeavor became an attempt to prove that CNN’s are as effective as possible in evaluating research variants with the lowest amount of carrier, for which the key aspect will be the ability to reduce energy costs in the manufacturing process of the final product. Other previous studies have taken into account that raspberry powders should be characterized by a regular shape of microparticles and contain the smallest possible intermolecular space [20]. In the current study results, the regular shape of these microparticles was also taken into account for comparison using LF-NMR and instrumental texture. Both in the previous studies and in the current one, the acquisition of a regular shape and an ordered interparticle structure is the key aspect in obtaining powders with a high quality index (homogeneity).

The current research examined how other carriers behave when extracting fruit powders. The focus was on the smallest and largest proportion of the carrier. As a result of the current and previous studies, it is confirmed that the recognition of regular microparticle shape and ordered intermolecular structure at 50% carrier content occurs with inulin and maltodextrin, among others [15]. When the ratio is reduced to 30% of the carrier content, it is confirmed to previous studies that identification of microparticle structure occurs best with inulin [12,20]. In the current study, milk whey proteins and microcrystalline cellulose were also successfully identified by CNN models as well as interpreted by textural parameters and the LF-NMR method.

In the study of texture analysis for the demonstrated carriers, statistically good solubility, hardness, cohesion, and viscosity of the obtained blackcurrant powders by spray drying are confirmed. At 30% carrier content in solution, maltodextrin and whey milk proteins determine the lowest Force_1 content underlying the linearity of the powders, as well as a longer T 1 relaxation time determining their homogeneity and microparticle ordering. In the case of 50% carrier in the blackcurrant solution, the lowest Force_1 content is determined for maltodextrin (MD50), inulin (IN50), milk whey protein (W50), and microstaline cellulose (C50). For comparison by the LF-NMR method, the shortest T 1 relaxation time was acquired by the C50 variant as one of the two blackcurrant powder variants (also F50). Evaluating the other textural parameters and discussing the results, it turns out that maltodextrin, inulin, milk whey proteins, and microcrystalline cellulose become quite effective carriers in determining the appropriate morphological structure.

A machine learning technique using computer vision has successfully determined which models have the highest quality index when recognizing the morphological structure of blackcurrant powders using image pixels. Conducting physical analyses using parameters such as texture, spin–lattice T 1 relaxation time using LF-NMR requires wider costs, more energy, and specialized training in their use. Therefore, it seems reasonable to use the latest artificial intelligence methods to assist in faster determination of the morphological structure of microparticles. CNNs were able to effectively identify the morphological structure of blackcurrant powders much faster and more efficiently than analytical methods. This translates into optimization of many unit processes in obtaining the final product. Machine learning is becoming an alternative solution to support the process of obtaining high-quality powders with desired properties. As a result of controlling the morphological structure of fruit powders, it becomes possible to optimize the spray drying process based on computer vision. Obtaining fruit powders with homogeneous microparticle shape and small inter-particle space in the spray drying process and evaluating them effectively using machine learning becomes a key criterion for maintaining the reproducibility of these results. This implies working on producing an intelligent powder obtaining system that controls the morphological structure of powders based on online computer vision.

5 Conclusions

In conclusion, according to the results obtained, it is possible to identify blackcurrant powder microparticles using CNNs and the Python language. The training CNNs gave the best results using carriers such as milk whey protein, inulin, and microcrystalline cellulose (designated W70-IN70 and C70-W70). The CNN models designated W70-IN70 and CNN C70-W70 achieved the lowest MSE values.

Currently, the aim is to improve a range of decision-making problems using advanced techniques by building machine learning models, data visualization, automation, and optimization of production processes using Python. Artificial intelligence is still evolving and the food industry can expect even further advances. Neural models can assist in reducing food waste, and AI-based robotics can revolutionize food processing and distribution. In the future, the experience will allow the implementation of online machine learning and deep learning solutions in the food industry as well. It is intended to implement convolutional network models in designing a homogeneous and ordered morphological structure of microparticles to obtain high-quality powders during spray drying (online). This will result in a final product with preserved nutritional values and their effective shelf life.

  1. Funding information: The work was financially supported by a grant from the National Science Centre (Poland) No. 2021/05/X/NZ9/01014 (PI: Krzysztof Przybył). The publication was co-financed by the Polish Minister of Science and Higher Education as part of the Strategy of the Poznan University of Life Sciences for 2024–2026 in the field of improving scientific research and development work in priority research areas.

  2. Author contributions: Conceptualization: Krzysztof Przybył; methodology: Krzysztof Przybył, Katarzyna Samborska, Hanna Maria Baranowska; software: Krzysztof Przybył; validation: Krzysztof Przybył; formal analysis: Krzysztof Przybył; investigation: Łukasz Masewicz, Aleksandra Jedlińska; resources: Krzysztof Przybył, Hanna Maria Baranowska, Przemysław Łukasz Kowalczewski; data curation: Krzysztof Przybył; writing – original draft preparation: Krzysztof Przybył; writing – review and editing: Krzysztof Przybył, Przemysław Łukasz Kowalczewski, Krzysztof Koszela, Katarzyna Samborska, Hanna Maria Baranowska, Łukasz Masewicz, Aleksandra Jedlińska; visualization: Krzysztof Przybył; supervision: Katarzyna Samborska, Hanna Maria Baranowska, Przemysław Łukasz Kowalczewski, Krzysztof Koszela. All authors have accepted responsibility for the entire content of this manuscript and approved its submission.

  3. Conflict of interest: The authors state no conflict of interest.

  4. Data availability statement: The data presented in this study are openly available in Repository for Open Data - RepOD at DOI: https://doi.org/10.18150/J6R7UY.

References

[1] Focker, M., E. D. van Asselt, B. J. A. Berendsen, M. G. M. van de Schans, S. P. J. van Leeuwen, S. M. Visser, et al. Review of food safety hazards in circular food systems in Europe. Food Research International, Vol. 158, 2022 Aug, id. 111505. https://linkinghub.elsevier.com/retrieve/pii/S0963996922005634.10.1016/j.foodres.2022.111505Search in Google Scholar PubMed

[2] Iranshahi, K., D. I. Onwude, A. Martynenko, and T. Defraeye. Dehydration mechanisms in electrohydrodynamic drying of plant-based foods. Food and Bioproducts Processing, Vol. 131, 2022 Jan, pp. 202–216. https://linkinghub.elsevier.com/retrieve/pii/S0960308521001838.10.1016/j.fbp.2021.11.009Search in Google Scholar

[3] Richter Reis, F., C. Marques, A. C. S. Moraes, and M. L. de, Masson. Trends in quality assessment and drying methods used for fruits and vegetables. Food Control, Vol. 142, 2022 Dec, id. 109254. https://linkinghub.elsevier.com/retrieve/pii/S0956713522004479.10.1016/j.foodcont.2022.109254Search in Google Scholar

[4] Xu, X., Y. Lu, B. Vogel-Heuser, and L. Wang. Industry 4.0 and Industry 5.0—inception, conception and perception. Journal of Manufacturing Systems, Vol. 61, 2021 Oct, pp. 530–535. https://linkinghub.elsevier.com/retrieve/pii/S0278612521002119.10.1016/j.jmsy.2021.10.006Search in Google Scholar

[5] Lingayat, A. B., V. P. Chandramohan, V. R. K. Raju, and V. Meda. A review on indirect type solar dryers for agricultural crops – dryer setup, its performance, energy storage and important highlights. Applied Energy, Vol. 258, 2020 Jan, id. 114005. https://linkinghub.elsevier.com/retrieve/pii/S0306261919316927.10.1016/j.apenergy.2019.114005Search in Google Scholar

[6] Raponi, F., R. Moscetti, D. Monarca, A. Colantoni, and R. Massantini. Monitoring and optimization of the process of drying fruits and vegetables using computer vision: a review. Sustainability, Vol. 9, No. 11, 2017 Nov, id. 2009. https://www.mdpi.com/2071-1050/9/11/2009.10.3390/su9112009Search in Google Scholar

[7] Gill, H. S., G. Murugesan, A. Mehbodniya, G. Sekhar Sajja, G. Gupta, and A. Bhatt. Fruit type classification using deep learning and feature fusion. Computers and Electronics in Agriculture, Vol. 211, 2023 Aug, id. 107990. https://linkinghub.elsevier.com/retrieve/pii/S0168169923003782.10.1016/j.compag.2023.107990Search in Google Scholar

[8] Massaro, A. Advanced control systems in industry 5.0 enabling process mining. Sensors, Vol. 22, No. 22, 2022 Nov, id. 8677. https://www.mdpi.com/1424-8220/22/22/8677.10.3390/s22228677Search in Google Scholar PubMed PubMed Central

[9] Min, W., Z. Wang, J. Yang, C. Liu, and S. Jiang. Vision-based fruit recognition via multi-scale attention CNN. Computers and Electronics in Agriculture, Vol. 210, 2023 Jul, id. 107911. https://linkinghub.elsevier.com/retrieve/pii/S0168169923002995.10.1016/j.compag.2023.107911Search in Google Scholar

[10] Zheng, M., S. Zhang, Y. Zhang, and B. Hu. Construct food safety traceability system for people’s health under the internet of things and big data. IEEE Access, Vol. 9, 2021, pp. 70571–70583. https://ieeexplore.ieee.org/document/9427143/.10.1109/ACCESS.2021.3078536Search in Google Scholar

[11] Boobier, T. Advanced analytics and AI: impact, implementation, and the future of work, John Wiley & Sons, Ltd, Hoboken, New Jersey, USA, 2018.10.1002/9781119390961Search in Google Scholar

[12] Przybył, K. and K. Koszela. Applications MLP and other methods in artificial intelligence of fruit and vegetable in convective and spray drying. Applied Sciences, Vol. 13, No. 5, 2023 Feb, id. 2965. https://www.mdpi.com/2076-3417/13/5/2965.10.3390/app13052965Search in Google Scholar

[13] Sun, Q., M. Zhang, and A. S. Mujumdar. Recent developments of artificial intelligence in drying of fresh food: a review. Critical Reviews in Food Science and Nutrition, Vol. 59, No. 14, 2019 Aug, pp. 2258–2275. https://www.tandfonline.com/doi/full/10.1080/10408398.2018.1446900.10.1080/10408398.2018.1446900Search in Google Scholar PubMed

[14] Jedlińska, A., K. Samborska, A. Wiktor, M. Balik, D. Derewiaka, A. Matwijczuk, et al. Spray drying of pure kiwiberry pulp in dehumidified air. Drying Technology, Vol. 40, No. 7, 2022 Jun, pp. 1421–1435. https://www.tandfonline.com/doi/full/10.1080/07373937.2020.1871006.10.1080/07373937.2020.1871006Search in Google Scholar

[15] Przybył, K., K. Koszela, F. Adamski, K. Samborska, K. Walkowiak, and M. Polarczyk. Deep and machine learning using SEM, FTIR, and texture analysis to detect polysaccharide in raspberry powders. Sensors, Vol. 21, No. 17, 2021 Aug, id. 5823. https://www.mdpi.com/1424-8220/21/17/5823.10.3390/s21175823Search in Google Scholar PubMed PubMed Central

[16] Samborska, K., A. Jedlińska, A. Wiktor, D. Derewiaka, R. Wołosiak, A. Matwijczuk, et al. The effect of low-temperature spray drying with dehumidified air on phenolic compounds, antioxidant activity, and aroma compounds of rapeseed honey powders. Food and Bioprocess Technology, Vol. 12, No. 6, 2019 Jun, pp. 919–932. http://link.springer.com/10.1007/s11947-019-02260-8.10.1007/s11947-019-02260-8Search in Google Scholar

[17] Feguš, U., U. Žigon, M. Petermann, and Ž. Knez. Effect of drying parameters on physiochemical and sensory properties of fruit powders processed by PGSS-, vacuum- and spray-drying. Acta Chimica Slovenica, Vol. 62, No. 2, 2015 Jun, pp. 479–487. https://journals.matheo.si/index.php/ACSi/article/view/969.10.17344/acsi.2014.969Search in Google Scholar PubMed

[18] Santos, D., A. C. Maurício, V. Sencadas, J. D. Santos, M. H. Fernandes, and P. S. Gomes. Spray drying: an overview. In Biomaterials – physics and chemistry – New Edition, InTech, London, UK, 2018. http://www.intechopen.com/books/biomaterials-physics-and-chemistry-new-edition/spray-drying-an-overview.10.5772/intechopen.72247Search in Google Scholar

[19] Shishir, M. R. I. and W. Chen. Trends of spray drying: a critical review on drying of fruit and vegetable juices. Trends in Food Science and Technology, Vol. 65, 2017 Jul, pp. 49–67. https://linkinghub.elsevier.com/retrieve/pii/S0924224416306355.10.1016/j.tifs.2017.05.006Search in Google Scholar

[20] Przybył, K., K. Samborska, K. Koszela, L. Masewicz, and T. Pawlak. Artificial neural networks in the evaluation of the influence of the type and content of carrier on selected quality parameters of spray dried raspberry powders. Measurement, Vol. 186, 2021 Dec, id. 110014. https://linkinghub.elsevier.com/retrieve/pii/S0263224121009416.10.1016/j.measurement.2021.110014Search in Google Scholar

[21] Jeżowski, P., J. Menzel, H. M. Baranowska, and P. Ł. Kowalczewski. Microwaved-assisted synthesis of starch-based biopolymer membranes for novel green electrochemical energy storage devices. Materials (Basel), Vol. 16, No. 22, 2023 Nov, id. 7111. https://www.mdpi.com/1996-1944/16/22/7111.10.3390/ma16227111Search in Google Scholar PubMed PubMed Central

[22] Luo, J., M. Li, Y. Zhang, M. Zheng, and C. Ming Ling. The low‐field NMR studies the change in cellular water in tilapia fillet tissue during different drying conditions. Food Science & Nutrition, Vol. 9, No. 5, 2021 May, pp. 2644–2657. https://onlinelibrary.wiley.com/doi/10.1002/fsn3.2221.10.1002/fsn3.2221Search in Google Scholar PubMed PubMed Central

[23] Nestle, N. NMR relaxometry study of cement hydration in the presence of different oxidic fine fraction materials. Solid State Nuclear Magnetic Resonance, Vol. 25, No. 1–3, 2004 Jan, pp. 80–83. https://linkinghub.elsevier.com/retrieve/pii/S0926204003000870.10.1016/j.ssnmr.2003.05.003Search in Google Scholar PubMed

[24] Walkowiak, K., K. Przybył, H. M. Baranowska, K. Koszela, Ł. Masewicz, and M. Piątek. The process of pasting and gelling modified potato starch with LF-NMR. Polymers (Basel), Vol. 14, No. 1, 2022 Jan, id. 184, https://www.mdpi.com/2073-4360/14/1/184.10.3390/polym14010184Search in Google Scholar PubMed PubMed Central

[25] Mao, Y., H. Li, Y. Wang, H. Wang, J. Shen, Y. Xu, et al. Rapid monitoring of tea plants under cold stress based on UAV multi-sensor data. Computers and Electronics in Agriculture, Vol. 213, 2023 Oct, id. 108176. https://linkinghub.elsevier.com/retrieve/pii/S0168169923005641.10.1016/j.compag.2023.108176Search in Google Scholar

[26] Sejnowski, T. J. The deep learning revolution, The MIT Press, Cambridge, MA, USA, 2018.10.7551/mitpress/11474.001.0001Search in Google Scholar

[27] Bremser, W. NMR macht Spaß! Experimental pulse NMR: a nuts and bolts approach. Von E. Fukushima/S. B. W. Roeder. Addison‐Wesley Publishing Company Inc. Reading, MA‐London‐Amsterdam‐Don Mills‐Ontario‐Sydney‐Tokyo 1981. XIV, 539 S., geb., $34,50. Nachrichten aus Chemie, Technik und Laboratorium, Vol. 30, No. 2, 1982 Feb, id. 144. https://onlinelibrary.wiley.com/doi/10.1002/nadc.19820300213.10.1002/nadc.19820300213Search in Google Scholar

[28] Fukushima, E. and S. B. W. Roeder. Experimental pulse NMR, CRC Press, Boca Raton, Florida, USA, 2018. https://www.taylorfrancis.com/books/9780429962417.Search in Google Scholar

[29] Węglarz, W. P. and H. Harańczyk. Two-dimensional analysis of the nuclear relaxation function in the time domain: the program CracSpin. Journal of Physics D: Applied Physics, Vol. 33, No. 15, 2000 Aug, pp. 1909–1920. http://stacks.iop.org/0022-3727/33/i=15/a=322?key=crossref.d9297399fd9aa25ff88d352dfa89b19b.10.1088/0022-3727/33/15/322Search in Google Scholar

[30] Lawless, C., D. J. Wilkinson, A. Young, S. G. Addinall, and D. A. Lydall. Colonyzer: automated quantification of micro-organism growth characteristics on solid agar. BMC Bioinformatics, Vol. 11, No. 1, 2010 Dec, id. 287. https://bmcbioinformatics.biomedcentral.com/articles/10.1186/1471-2105-11-287.10.1186/1471-2105-11-287Search in Google Scholar PubMed PubMed Central

[31] Guo, X., Y. Ge, F. Liu, and J. Yang. Identification of maize and wheat seedlings and weeds based on deep learning. Frontiers in Earth Science, Vol. 11, 2023 Feb, id. 1146558. https://www.frontiersin.org/articles/10.3389/feart.2023.1146558/full.10.3389/feart.2023.1146558Search in Google Scholar

[32] Krizhevsky, A., I. Sutskever, and G. E. Hinton. ImageNet classification with deep convolutional neural networks. Communications of the ACM, Vol. 60, No. 6, 2017 May, pp. 84–90. https://dl.acm.org/doi/10.1145/3065386.10.1145/3065386Search in Google Scholar

[33] Rauf, H. T., M. I. U. Lali, S. Zahoor, S. Z. H. Shah, A. U. Rehman, and S. A. C. Bukhari. Visual features based automated identification of fish species using deep convolutional neural networks. Computers and Electronics in Agriculture, Vol. 167, 2019 Dec, id. 105075. https://linkinghub.elsevier.com/retrieve/pii/S0168169919313523.10.1016/j.compag.2019.105075Search in Google Scholar

[34] Chollet, F. Deep learning with python, Manning Publications Co., Cham, Switzerland, 2017.Search in Google Scholar

[35] Yang, J. and G. Yang. Modified convolutional neural network based on dropout and the stochastic gradient descent optimizer. Algorithms, Vol. 11, No. 3, 2018 Mar, id. 28. https://www.mdpi.com/1999-4893/11/3/28.10.3390/a11030028Search in Google Scholar

[36] Kingma, D. P. and J. Ba. Adam: a method for stochastic optimization, 2017 Jan. http://arxiv.org/abs/1412.6980.Search in Google Scholar

[37] Zou, F., L. Shen, Z. Jie, W. Zhang, and W. Liu. A sufficient condition for convergences of adam and RMSProp. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), IEEE, 2019, pp. 11119–11127. https://ieeexplore.ieee.org/document/8954145/.10.1109/CVPR.2019.01138Search in Google Scholar

[38] Ogundokun, R. O., R. Maskeliunas, S. Misra, and R. Damaševičius. Improved CNN based on batch normalization and adam optimizer. In: Computational Science and Its Applications – ICCSA 2022 Workshops. ICCSA 2022. Lecture Notes in Computer Science, Vol. 13381, Springer, Cham, 2022, pp. 593–604. https://doi.org/10.1007/978-3-031-10548-7_43.10.1007/978-3-031-10548-7_43Search in Google Scholar

[39] Bashir, D., G. D. Montañez, S. Sehra, P. S. Segura, and J. Lauw. An information-theoretic perspective on overfitting and underfitting. In: AI 2020: Advances in Artificial Intelligence, AI 2020, Lecture Notes in Computer Science, Vol 12576. Springer, Cham, 2020, pp. 347–358. https://doi.org/10.1007/978-3-030-64984-5_27.10.1007/978-3-030-64984-5_27Search in Google Scholar

[40] Kuti, T., A. Hegyi, and S. Kemény. Analysis of sensory data of different food products by ANOVA. Chemometrics and Intelligent Laboratory Systems, Vol. 72, No. 2, 2004 Jul, pp. 253–257. https://linkinghub.elsevier.com/retrieve/pii/S0169743904000437.10.1016/j.chemolab.2004.01.020Search in Google Scholar

[41] Rust, A., F. Marini, M. Allsopp, P. J. Williams, and M. Manley. Application of ANOVA-simultaneous component analysis to quantify and characterise effects of age, temperature, syrup adulteration and irradiation on near-infrared (NIR) spectral data of honey. Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy, Vol. 253, 2021 May, id. 119546, https://linkinghub.elsevier.com/retrieve/pii/S1386142521001220.10.1016/j.saa.2021.119546Search in Google Scholar PubMed

[42] Guo, Y., Y. Zhou, and Z. Zhang. Fault diagnosis of multi-channel data by the CNN with the multilinear principal component analysis. Measurement, Vol. 171, 2021 Feb, id. 108513, https://linkinghub.elsevier.com/retrieve/pii/S026322412031040X.10.1016/j.measurement.2020.108513Search in Google Scholar

[43] Ahmad, S., I. Pasha, M. Saeed, and M. Shahid. Principal component analysis and correlation studies of spring wheats in relation to cookie making quality. International Journal of Food Properties, Vol. 20, No. 10, 2017 Oct, pp. 2299–2313. https://www.tandfonline.com/doi/full/10.1080/10942912.2016.1236273.10.1080/10942912.2016.1236273Search in Google Scholar

[44] Jolliffe, I. T. Principal component analysis, Springer-Verlag, New York, 2002. (Springer Series in Statistics). http://link.springer.com/10.1007/b98835.Search in Google Scholar

[45] Flowers, B., K.-T. Huang, and G. O. Aldana. Analysis of the habitat fragmentation of ecosystems in belize using landscape metrics. Sustainability, Vol. 12, No. 7, 2020 Apr, id. 3024. https://www.mdpi.com/2071-1050/12/7/3024.10.3390/su12073024Search in Google Scholar

[46] Sudha Mishra, S. and A. K. Das Mohapatra. Weavers’ perception towards sustainability of Sambalpuri handloom: a Tukey’s HSD test analysis. Materials Today: Proceedings, Vol. 51, 2022, pp. 217–227. https://linkinghub.elsevier.com/retrieve/pii/S2214785321038335.10.1016/j.matpr.2021.05.242Search in Google Scholar

[47] Frochte, J. Python, NumPy, SciPy und Matplotlib – in a nutshell. In Maschinelles Lernen, Carl Hanser Verlag GmbH & Co. KG, München, 2020, pp. 38–77. https://www.hanser-elibrary.com/doi/10.3139/9783446463554.003.10.3139/9783446463554.003Search in Google Scholar

[48] van der Walt, S., J. L. Schönberger, J. Nunez-Iglesias, F. Boulogne, J. D. Warner, N. Yager, et al. Scikit-image: image processing in Python. PeerJ, Vol. 2, 2014 Jun, id. e453. https://peerj.com/articles/453.10.7717/peerj.453Search in Google Scholar PubMed PubMed Central

[49] Di Monaco, R., S. Cavella, and P. Masi. Predicting sensory cohesiveness, hardness and springiness of solid foods from instrumental measurements. Journal of Texture Studies, Vol. 39, No. 2, 2008 Apr, pp. 129–149. https://onlinelibrary.wiley.com/doi/10.1111/j.1745-4603.2008.00134.x.10.1111/j.1745-4603.2008.00134.xSearch in Google Scholar

[50] Rosenthal, A. J. and P. Thompson. What is cohesiveness? A linguistic exploration of the food texture testing literature. Journal of Texture Studies, Vol. 52, No. 3, 2021 Jun, pp. 294–302. https://onlinelibrary.wiley.com/doi/10.1111/jtxs.12586.10.1111/jtxs.12586Search in Google Scholar PubMed

[51] Tobin, A. B., P. Heunemann, J. Wemmer, J. R. Stokes, T. Nicholson, E. J. Windhab, et al. Cohesiveness and flowability of particulated solid and semi-solid food systems. Food & Function, Vol. 8, No. 10, 2017, pp. 3647–3653. https://xlink.rsc.org/?DOI=C7FO00715A.10.1039/C7FO00715ASearch in Google Scholar

[52] Lim, H. L., K. P. Hapgood, and B. Haig. Understanding and preventing agglomeration in a filter drying process. Powder Technology, Vol. 300, 2016 Oct, pp. 146–156. https://linkinghub.elsevier.com/retrieve/pii/S0032591016300961.10.1016/j.powtec.2016.03.003Search in Google Scholar

[53] Różańska, M. B., J. Zembrzuska, P. Rychlewski, M. Kidoń, Ł. Masewicz, S. Mildner-Szkudlarz, et al. Exploring the impact of dietary fiber enrichment on molecular water properties and indicators of Maillard reaction (furosine, Nε-carboxymethyllysine, and Nε-carboxyethyllysine) in model gluten-free bread. Food Chemistry, Vol. 491, 2025 Nov, id. 145194. https://linkinghub.elsevier.com/retrieve/pii/S0308814625024458.10.1016/j.foodchem.2025.145194Search in Google Scholar PubMed

[54] Przybył, K., J. Gawałek, K. Koszela, J. Wawrzyniak, and L. Gierz. Artificial neural networks and electron microscopy to evaluate the quality of fruit and vegetable spray-dried powders. Case study: strawberry powder. Computers and Electronics in Agriculture, Vol. 155, 2018 Dec, pp. 314–323. https://linkinghub.elsevier.com/retrieve/pii/S0168169918303466.10.1016/j.compag.2018.10.033Search in Google Scholar
Received: 2024-03-12
Revised: 2025-07-07
Accepted: 2025-07-08
Published Online: 2025-08-22

© 2025 the author(s), published by De Gruyter

This work is licensed under the Creative Commons Attribution 4.0 International License.

Articles in the same Issue

  1. Review Articles
  2. Utilization of steel slag in concrete: A review on durability and microstructure analysis
  3. Technical development of modified emulsion asphalt: A review on the preparation, performance, and applications
  4. Recent developments in ultrasonic welding of similar and dissimilar joints of carbon fiber reinforcement thermoplastics with and without interlayer: A state-of-the-art review
  5. Unveiling the crucial factors and coating mitigation of solid particle erosion in steam turbine blade failures: A review
  6. From magnesium oxide, magnesium oxide concrete to magnesium oxide concrete dams
  7. Properties and potential applications of polymer composites containing secondary fillers
  8. A scientometric review on the utilization of copper slag as a substitute constituent of ordinary Portland cement concrete
  9. Advancement of additive manufacturing technology in the development of personalized in vivo and in vitro prosthetic implants
  10. Recent advance of MOFs in Fenton-like reaction
  11. A review of defect formation, detection, and effect on mechanical properties of three-dimensional braided composites
  12. Non-conventional approaches to producing biochars for environmental and energy applications
  13. Review of the development and application of aluminum alloys in the nuclear industry
  14. Advances in the development and characterization of combustible cartridge cases and propellants: Preparation, performance, and future prospects
  15. Recent trends in rubberized and non-rubberized ultra-high performance geopolymer concrete for sustainable construction: A review
  16. Cement-based materials for radiative cooling: Potential, material and structural design, and future prospects
  17. A comprehensive review: The impact of recycling polypropylene fiber on lightweight concrete performance
  18. Research Articles
  19. Investigation of the corrosion performance of HVOF-sprayed WC-CoCr coatings applied on offshore hydraulic equipment
  20. A systematic review of metakaolin-based alkali-activated and geopolymer concrete: A step toward green concrete
  21. Evaluation of color matching of three single-shade composites employing simulated 3D printed cavities with different thicknesses using CIELAB and CIEDE2000 color difference formulae
  22. Novel approaches in prediction of tensile strain capacity of engineered cementitious composites using interpretable approaches
  23. Effect of TiB2 particles on the compressive, hardness, and water absorption responses of Kulkual fiber-reinforced epoxy composites
  24. Analyzing the compressive strength, eco-strength, and cost–strength ratio of agro-waste-derived concrete using advanced machine learning methods
  25. Tensile behavior evaluation of two-stage concrete using an innovative model optimization approach
  26. Tailoring the mechanical and degradation properties of 3DP PLA/PCL scaffolds for biomedical applications
  27. Optimizing compressive strength prediction in glass powder-modified concrete: A comprehensive study on silicon dioxide and calcium oxide influence across varied sample dimensions and strength ranges
  28. Experimental study on solid particle erosion of protective aircraft coatings at different impact angles
  29. Compatibility between polyurea resin modifier and asphalt binder based on segregation and rheological parameters
  30. Fe-containing nominal wollastonite (CaSiO3)–Na2O glass-ceramic: Characterization and biocompatibility
  31. Relevance of pore network connectivity in tannin-derived carbons for rapid detection of BTEX traces in indoor air
  32. A life cycle and environmental impact analysis of sustainable concrete incorporating date palm ash and eggshell powder as supplementary cementitious materials
  33. Eco-friendly utilisation of agricultural waste: Assessing mixture performance and physical properties of asphalt modified with peanut husk ash using response surface methodology
  34. Benefits and limitations of N2 addition with Ar as shielding gas on microstructure change and their effect on hardness and corrosion resistance of duplex stainless steel weldments
  35. Effect of selective laser sintering processing parameters on the mechanical properties of peanut shell powder/polyether sulfone composite
  36. Impact and mechanism of improving the UV aging resistance performance of modified asphalt binder
  37. AI-based prediction for the strength, cost, and sustainability of eggshell and date palm ash-blended concrete
  38. Investigating the sulfonated ZnO–PVA membrane for improved MFC performance
  39. Strontium coupling with sulphur in mouse bone apatites
  40. Transforming waste into value: Advancing sustainable construction materials with treated plastic waste and foundry sand in lightweight foamed concrete for a greener future
  41. Evaluating the use of recycled sawdust in porous foam mortar for improved performance
  42. Improvement and predictive modeling of the mechanical performance of waste fire clay blended concrete
  43. Polyvinyl alcohol/alginate/gelatin hydrogel-based CaSiO3 designed for accelerating wound healing
  44. Research on assembly stress and deformation of thin-walled composite material power cabin fairings
  45. Effect of volcanic pumice powder on the properties of fiber-reinforced cement mortars in aggressive environments
  46. Analyzing the compressive performance of lightweight foamcrete and parameter interdependencies using machine intelligence strategies
  47. Selected materials techniques for evaluation of attributes of sourdough bread with Kombucha SCOBY
  48. Establishing strength prediction models for low-carbon rubberized cementitious mortar using advanced AI tools
  49. Investigating the strength performance of 3D printed fiber-reinforced concrete using applicable predictive models
  50. An eco-friendly synthesis of ZnO nanoparticles with jamun seed extract and their multi-applications
  51. The application of convolutional neural networks, LF-NMR, and texture for microparticle analysis in assessing the quality of fruit powders: Case study – blackcurrant powders
  52. Study of feasibility of using copper mining tailings in mortar production
  53. Shear and flexural performance of reinforced concrete beams with recycled concrete aggregates
  54. Advancing GGBS geopolymer concrete with nano-alumina: A study on strength and durability in aggressive environments
  55. 10.1515/rams-2025-0143
  56. 10.1515/rams-2025-0144
  57. 10.1515/rams-2025-0130
  58. Special Issue on Recent Advancement in Low-carbon Cement-based Materials - Part II
  59. Investigating the effect of locally available volcanic ash on mechanical and microstructure properties of concrete
  60. Flexural performance evaluation using computational tools for plastic-derived mortar modified with blends of industrial waste powders
  61. Foamed geopolymers as low carbon materials for fire-resistant and lightweight applications in construction: A review
  62. Autogenous shrinkage of cementitious composites incorporating red mud
  63. Special Issue on AI-Driven Advances for Nano-Enhanced Sustainable Construction Materials
  64. Advanced explainable models for strength evaluation of self-compacting concrete modified with supplementary glass and marble powders
  65. Special Issue on Advanced Materials for Energy Storage and Conversion
  66. 10.1515/rams-2025-0131
https://doi.org/10.1515/rams-2025-0132
Keywords for this article
convolutional neural networks; low-field nuclear magnetic resonance; texture analysis; scanning electron microscopy; fruit powders; blackcurrant
Creative Commons
BY 4.0

Articles in the same Issue

  1. Review Articles
  2. Utilization of steel slag in concrete: A review on durability and microstructure analysis
  3. Technical development of modified emulsion asphalt: A review on the preparation, performance, and applications
  4. Recent developments in ultrasonic welding of similar and dissimilar joints of carbon fiber reinforcement thermoplastics with and without interlayer: A state-of-the-art review
  5. Unveiling the crucial factors and coating mitigation of solid particle erosion in steam turbine blade failures: A review
  6. From magnesium oxide, magnesium oxide concrete to magnesium oxide concrete dams
  7. Properties and potential applications of polymer composites containing secondary fillers
  8. A scientometric review on the utilization of copper slag as a substitute constituent of ordinary Portland cement concrete
  9. Advancement of additive manufacturing technology in the development of personalized in vivo and in vitro prosthetic implants
  10. Recent advance of MOFs in Fenton-like reaction
  11. A review of defect formation, detection, and effect on mechanical properties of three-dimensional braided composites
  12. Non-conventional approaches to producing biochars for environmental and energy applications
  13. Review of the development and application of aluminum alloys in the nuclear industry
  14. Advances in the development and characterization of combustible cartridge cases and propellants: Preparation, performance, and future prospects
  15. Recent trends in rubberized and non-rubberized ultra-high performance geopolymer concrete for sustainable construction: A review
  16. Cement-based materials for radiative cooling: Potential, material and structural design, and future prospects
  17. A comprehensive review: The impact of recycling polypropylene fiber on lightweight concrete performance
  18. Research Articles
  19. Investigation of the corrosion performance of HVOF-sprayed WC-CoCr coatings applied on offshore hydraulic equipment
  20. A systematic review of metakaolin-based alkali-activated and geopolymer concrete: A step toward green concrete
  21. Evaluation of color matching of three single-shade composites employing simulated 3D printed cavities with different thicknesses using CIELAB and CIEDE2000 color difference formulae
  22. Novel approaches in prediction of tensile strain capacity of engineered cementitious composites using interpretable approaches
  23. Effect of TiB2 particles on the compressive, hardness, and water absorption responses of Kulkual fiber-reinforced epoxy composites
  24. Analyzing the compressive strength, eco-strength, and cost–strength ratio of agro-waste-derived concrete using advanced machine learning methods
  25. Tensile behavior evaluation of two-stage concrete using an innovative model optimization approach
  26. Tailoring the mechanical and degradation properties of 3DP PLA/PCL scaffolds for biomedical applications
  27. Optimizing compressive strength prediction in glass powder-modified concrete: A comprehensive study on silicon dioxide and calcium oxide influence across varied sample dimensions and strength ranges
  28. Experimental study on solid particle erosion of protective aircraft coatings at different impact angles
  29. Compatibility between polyurea resin modifier and asphalt binder based on segregation and rheological parameters
  30. Fe-containing nominal wollastonite (CaSiO3)–Na2O glass-ceramic: Characterization and biocompatibility
  31. Relevance of pore network connectivity in tannin-derived carbons for rapid detection of BTEX traces in indoor air
  32. A life cycle and environmental impact analysis of sustainable concrete incorporating date palm ash and eggshell powder as supplementary cementitious materials
  33. Eco-friendly utilisation of agricultural waste: Assessing mixture performance and physical properties of asphalt modified with peanut husk ash using response surface methodology
  34. Benefits and limitations of N2 addition with Ar as shielding gas on microstructure change and their effect on hardness and corrosion resistance of duplex stainless steel weldments
  35. Effect of selective laser sintering processing parameters on the mechanical properties of peanut shell powder/polyether sulfone composite
  36. Impact and mechanism of improving the UV aging resistance performance of modified asphalt binder
  37. AI-based prediction for the strength, cost, and sustainability of eggshell and date palm ash-blended concrete
  38. Investigating the sulfonated ZnO–PVA membrane for improved MFC performance
  39. Strontium coupling with sulphur in mouse bone apatites
  40. Transforming waste into value: Advancing sustainable construction materials with treated plastic waste and foundry sand in lightweight foamed concrete for a greener future
  41. Evaluating the use of recycled sawdust in porous foam mortar for improved performance
  42. Improvement and predictive modeling of the mechanical performance of waste fire clay blended concrete
  43. Polyvinyl alcohol/alginate/gelatin hydrogel-based CaSiO3 designed for accelerating wound healing
  44. Research on assembly stress and deformation of thin-walled composite material power cabin fairings
  45. Effect of volcanic pumice powder on the properties of fiber-reinforced cement mortars in aggressive environments
  46. Analyzing the compressive performance of lightweight foamcrete and parameter interdependencies using machine intelligence strategies
  47. Selected materials techniques for evaluation of attributes of sourdough bread with Kombucha SCOBY
  48. Establishing strength prediction models for low-carbon rubberized cementitious mortar using advanced AI tools
  49. Investigating the strength performance of 3D printed fiber-reinforced concrete using applicable predictive models
  50. An eco-friendly synthesis of ZnO nanoparticles with jamun seed extract and their multi-applications
  51. The application of convolutional neural networks, LF-NMR, and texture for microparticle analysis in assessing the quality of fruit powders: Case study – blackcurrant powders
  52. Study of feasibility of using copper mining tailings in mortar production
  53. Shear and flexural performance of reinforced concrete beams with recycled concrete aggregates
  54. Advancing GGBS geopolymer concrete with nano-alumina: A study on strength and durability in aggressive environments
  55. 10.1515/rams-2025-0143
  56. 10.1515/rams-2025-0144
  57. 10.1515/rams-2025-0130
  58. Special Issue on Recent Advancement in Low-carbon Cement-based Materials - Part II
  59. Investigating the effect of locally available volcanic ash on mechanical and microstructure properties of concrete
  60. Flexural performance evaluation using computational tools for plastic-derived mortar modified with blends of industrial waste powders
  61. Foamed geopolymers as low carbon materials for fire-resistant and lightweight applications in construction: A review
  62. Autogenous shrinkage of cementitious composites incorporating red mud
  63. Special Issue on AI-Driven Advances for Nano-Enhanced Sustainable Construction Materials
  64. Advanced explainable models for strength evaluation of self-compacting concrete modified with supplementary glass and marble powders
  65. Special Issue on Advanced Materials for Energy Storage and Conversion
  66. 10.1515/rams-2025-0131
Sign up now to receive a 20% welcome discount
Institutional Access
How does access work?
Have an idea on how to improve our website?
Please write us.
© 2025 De Gruyter Brill
Downloaded on 18.9.2025 from https://www.degruyterbrill.com/document/doi/10.1515/rams-2025-0132/html
Scroll to top button