CIEDE2000-based Wine Color Analysis Using Smartphones in Unconstrained Environments

© 2025 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. DOI: 10.1109/EEITE65381.2025.11166185 CIEDE2000-based Wine Color Analysis Using Smartphones in Unconstrained Environments Christina Mastralexi1,2 , Bruno Scholles Soares Dias1,2 , Clara Coelho1,3, Pedro Carvalho1,3 1Center of Telecommunications and Multimedia, INESC TEC, Porto, Portugal 2Faculty of Engineering, University of Porto (FEUP), Porto, Portugal 3ISEP, Polytechnic of Porto, Porto, Portugal Emails: [email protected], [email protected], [email protected], pedro.m.carv[email protected] Abstract—Wine color offers key insights into composition, origin, and aging, serving as a crucial indicator for classification and quality assessment. Traditional methods, based on sensory evaluation and spectrophotometry, are often environmentally sensitive and expensive. This study explores digital image analysis and perceptual color similarity metrics to achieve coarse wine classification in unconstrained environments using smartphones. Through machine learning, we identify color centroids for red, white, and ros´ e wines, employing threshold-based, Support Vector Machine (SVM), and closest-centroid approaches. By leveraging the CIEDE2000 metric alongside other distance-based techniques, our SVM model achieves 98% classification accuracy (F1-score = 0.979). This demonstrates a cost-effective, accessible alternative to invasive laboratory methods for wine color analysis, even under diverse, unconstrained conditions. Index Terms—CIEDE2000, image processing, unconstrained environments, wine classification, wine color analysis I. INTRODUCTION Wine color conveys key information about quality, origin, and aging, shaping consumer expectations of aroma and taste experience [1, 2]. Beyond aesthetics, wine color reflects its chemical composition, quality, and production process, influencing both appearance and sensory profile [3, 4]. Traditional wine analysis methods (sensory evaluation, spectrophotometry) are accurate yet expensive and less accessible to small-scale winemakers and consumers [1]. To address these challenges, digital image analysis offers a promising alternative, leveraging advancements in color metrics and computational techniques [5]. The CIELAB color space, recommended by the International Organization of Vine and Wine (OIV), correlates visually perceived colors with measured parameters [6]. However, its non-uniformities led to the CIEDE2000 formula, improving subtle color differentiation [5, 7]. Integrating digital imaging with perceptual metrics like CIEDE2000 enables wine color analysis in diverse settings, benefiting visually impaired users and everyday consumers who depend on accessible technological solutions for decisionmaking. To explore a more flexible and accessible approach to wine color analysis, we employed a combination of digital image processing and color difference metrics to classify wines based on their visual characteristics, and explore a flexible, cost-effective and accessible alternative that can be applied in diverse settings. Digital image-based techniques provide a promising approach for pixel-level color analysis for food and beverage quality, overcoming the limitations of traditional methods [2, 4]. In our work, wine images were manually segmented to isolate relevant areas, ensuring accurate color assessment [1]. The resulting color samples were then analyzed in both RGB and LAB color spaces, which allowed for a comprehensive understanding of wine color characteristics. Different color difference metrics, including Euclidean, Manhattan, CIE94, and CIEDE2000 distance, have been used to assess perceptual color differences relative to cluster centroids obtained via k-means clustering [7, 8]. The use of these metrics enables the identification of representative colors for each wine type and minimizes the need for comparison with a broad color palette. The CIEDE2000 metric, in particular, provided a more consistent measure of perceived differences, making it well-suited for distinguishing among similar color tones in wine samples [5]. The primary contributions of this study include: •A dataset of wine images, segmented and annotated. •The application of k-means clustering in conjunction with color difference metrics to define representative color centroids for red, white, and ros´ e wines. •A decision-making framework for wine classification based on multiple color distance metrics. Figure 1 presents an overview of the proposed pipeline for wine parameterization, from data acquisition and region-ofinterest extraction to color reference identification and threshold determination. Fig. 1. Pipeline illustrating the overall steps for the wine parameterization process using image analysis and color metrics. II. RELATED WORK Color similarity measurement is essential in digital imaging, particularly for accurately capturing perceptual differences. However, traditional Euclidean distance in RGB is limited due to its non-uniformity, while alternative metrics like ChiSquare and Chebyshev Distance are effective in classification and outlier detection. Likewise, traditional Euclidean and nonEuclidean metrics are often utilized in color spaces like CIELAB for small color differences. For large color differences, hybrid metrics like HyAB, which combine Euclidean and city-block metrics, offer enhanced effectiveness by separately addressing hue, chroma, and lightness [7]. The choice of metric depends on the application and the nature of the color data being analyzed. Perceptually uniform color spaces, notably CIELAB, have emerged as the standard for perceptual color analysis [5, 9]. The CIE76 formula was initially introduced but had limitations with high-chroma colors. To address this, CIE94 introduced weighting factors for lightness, chroma, and hue, enhancing perceptual differentiation [9]. Later, CIEDE2000 improved further upon CIE94 by fine-tuning these factors, and it has since become the preferred standard in color measurement [5]. Pathare et al. [10] and Mogol et al. [11] have highlighted how CIEDE2000 can be used for food and beverage quality, where color perception is critical. More recent advancements have introduced efficient CIEDE2000-based implementations in order to reduce computational demands [8, 12]. In the context of wine, color acts as a key indicator of quality, origin, aging, and authenticity, primarily influenced by phenolic compounds such as anthocyanins and tannins. Bright hues in young red wines gradually darken over time due to anthocyanin degradation and stable pigment formation [13, 14]. Traditional methods, like the Glories method, assess color intensity and hue via absorbance at specific wavelengths [1]. Although spectrophotometric techniques provide detailed and accurate color analysis, they necessitate costly equipment, laboratory conditions, and specialized personnel, thus limiting accessibility [15]. To overcome these constraints, digital imaging has enabled non-invasive, cost-effective methods for wine color analysis, utilizing color spaces like RGB, HSI, and CIELAB to assess authenticity and quality while minimizing reliance on labbased techniques [16]. Digital imaging has been successfully employed for wine authentication, adulteration detection, and improving consistency in color evaluation through standardized visual and photographic techniques [17, 18]. While offering flexibility and cost advantages, accuracy can be affected by lighting variability and human perception. III. DATASET ACQUISITION AND REPRESENTATION A. Dataset Preparation This study focuses on red, white, and ros´ e wines due to their distinct phenolic compositions, quality impact, and consumer relevance [19]. Established analytical methods for these wines are well-documented, while methodologies for other wine types remain less developed due to their smaller market presence [15]. For this study, we used a selection of wine images collected through two primary sources: (1) 75 images acquired using a custom-developed image acquisition application, and (2) a subset of 274 images selected from an online publicly available dataset [20]. Existing wine datasets primarily focus on chemical analysis and multispectral imaging rather than image-based classification. Examples include the UCI Wine Dataset [21] and Vinho Verde Dataset [22], which emphasize chemical features. Each image was manually segmented to isolate the wine’s region of interest (ROI) and minimize external influences and non-wine elements. Extracted RGB color values were stored with corresponding wine labels, as RGB is the standard format for consumer devices. However, LAB, derived from RGB, is more effective for color analysis by separating intensity from chromatic information, so both color spaces were used in our methodology. The low-resolution images, sourced from the online dataset, were upscaled using the super-resolution Real-ESRGAN (Real-Enhanced Super-Resolution Generative Adversarial Network) model [23] to ensure consistent region-of-interest (ROI) extraction. This is a deep learning-based model designed to upscale images while preserving fine details and minimizing artifacts. Before applying super-resolution, each image was evaluated based on its total pixel count. Images exceeding 1M pixels were deemed sufficiently detailed and retained without modification. Conversely, images below this threshold were upscaled using Real-ESRGAN, with a dynamically determined scaling factor: images smaller than 250kpixels (500 ×500) were upscaled by a factor of 5, whereas those between 250k and 1Mpixels were upscaled by a factor of 3. Finally, the resulting composition is shown in Table I, with the final resolution interval for the entire dataset ranging from a minimum of 1270 ×800 to a maximum of 3072 ×4096. TABLE I WINE COLOR CLASS COMPOSITION IN CUSTOM AND PUBLIC DATASET. Red White Ros´ e Custom Dataset 28 31 16 Public Dataset [20] 69 95 110 Total 97 126 126 B. Pixel Sampling Strategy and Data Distribution To ensure a comprehensive and unbiased color analysis, a structured sampling methodology is implemented. The principal challenge stems from the variability in image dimensions, which directly impacts the number of available pixels per sample. To mitigate this issue, two distinct sampling strategies are employed: •Full-Sample Analysis: Random selection of a fixed number of pixels, npixels, from all available pixels within an image. By sampling across the entire image, this method provides a holistic representation of the color distribution while maintaining a controlled sample size. Nonetheless, variations in image content and illumination conditions may introduce certain levels of bias. •Central Region Sampling (Radius-Based Sampling): To standardize the number of sampled pixels across different images, a circular region centered within each ROI is extracted. Given a predefined number of pixels, npixels, the sampling radius ris computed as follows: r=pnpixels ·Aimg/π min(H, W)(1) where Aimg represents the total image area, and Hand Wdenote the image height and width, respectively. This technique ensures that the sampled data remains consistent across different images, thereby minimizing potential sources of bias. IV. COLOR CHARACTERIZATION AND CENTROID ANALYSIS A. Reference Colors In order to identify representative reference colors for each type of wine, k-means clustering was used [24]. The goal of this clustering approach was to determine representative colors for the wine categories under study (red, white, and ros´ e), thus reducing computational complexity. The clustering analysis involved multiple experiments, varying sampling sizes, radius ratios, and methods for calculating central tendencies. Table II summarizes the resulting centroid coordinates and clustering accuracy scores, which evaluate how well each sampling strategy models the distribution of wine colors. B. Centroid Analysis The stability and representativeness of centroid coordinates were evaluated across multiple experiments to determine the most consistent cluster centers, as seen in Table II. Since centroids act as reference points for color classification, their reliability across different sampling strategies is essential for ensuring precise clustering and accurate predictions. •Red wine centroid: The RGB centroid for red wine remains highly stable across experiments. Similarly, LAB centroids exhibit minimal variation, with L values between 9 and 11 and only minor shifts in the A and B channels. This consistency indicates that red wine clustering is less affected by sampling variations. •White and ros´ e wine centroid: White and ros´ e wine centroids display greater fluctuation, particularly in the blue and green channels, indicating a broader hue range. The LAB coordinates confirm this variability, with L values for white wine ranging from 70 to 71 and for ros´ e wine from 51 to 52. These findings suggest that these categories are more sensitive to sampling differences. As shown in Table II, Experiment 3 delivers the most stable and accurate centroids for all three wine types. These coordinates align with the highest classification accuracies (up to 0.7542 in RGB and 0.8966 in LAB), making them the best representatives of each color category. Figure 2 compares k-means clustering results in LAB and RGB spaces. LAB clustering achieves better separation, especially for red wine, aligning closely with true labels. In contrast, RGB clustering shows more overlap between ros´ e and white wines, indicating LAB provides a more distinct representation for classification. Fig. 2. Example of K-Means clustering results for LAB and RGB sampled data, using 2,000 pixel samples per image in a radius of 0.5. This comparison illustrates the predicted cluster colors versus the true labels. V. COLOR-BASED WINE CLASSIFICATION This stage focuses on classifying wine samples based on their color attributes by systematically extracting key features from digital images, computing advanced distance metrics to centroids, and classify the sample. A. Decision Framework for Wine Classification Wine classification requires transitioning from pixel-level analysis, where each pixel is assessed individually, to an aggregated decision that evaluates the entire region of interest.For each extracted ROI, all pixels are compared against reference centroids using distance metrics. The classification is then determined based on the three following strategies: •Closest-Centroid Classification: The median distance of all pixels to each centroid is computed, and the sample is classified into the wine type with the smallest median distance. •Threshold-Based Classification: A wine sample is classified on a pixel-by-pixel basis, where each pixel’s distance to predefined centroids is measured and thresholded. Consequently, the final classification of the wine image is determined by majority vote. •SVM-based Classification: Each wine sample is classified at the pixel level. For each pixel, its distances to the reference centroids are computed using a specific colorbased metric. These three distances serve as input features for an SVM classifier, which determines the final wine classification based on majority voting across all pixels in an image. TABLE II CENTROID COORDINATES AND ACCURACY SCORES FOR DIFFERENT SAMPLING STRATEGIES IN WINE COLOR CLUSTERING. Experiment ID Radius Ratio Number of Pixels RGB Centroid Coordinates LAB Centroid Coordinates RGB Accuracy Score LAB Accuracy Score Red White Ros´ e Red White Ros´ e 1 Full Image 5000 (39, 22, 18) (220, 192, 141) (178, 100, 53) (11, 6, 4) (71, 0, 31) (51, 39, 42) 0.7262 0.8606 2 0.25 2000 (37, 22, 18) (218, 190, 131) (183, 97, 46) (10, 5, 4) (70, -1, 33) (52, 41, 45) 0.7498 0.8932 3 0.50 2000 (35, 20, 17) (218, 188, 129) (181, 97, 46) (9, 5, 4) (70, 0, 34) (52, 40, 45) 0.7542 0.8966 4 0.75 2000 (35, 20, 17) (219, 190, 132) (181, 97, 47) (9, 5, 4) (70, 0, 33) (52, 40, 44) 0.7533 0.8918 B. Distance Metrics for Color Comparison To classify wine samples, their color profiles are compared with centroids representing red, white, and ros´ e wines in both RGB and LAB color spaces. The following metrics are widely employed to quantify color similarity, providing distinct approaches to assess both perceptual and numerical differences in color. Euclidean Distance: The Euclidean distance quantifies the direct geometric separation between two points in a given color space. In the LAB space, this measure corresponds to ∆E76. Manhattan Distance: The Manhattan distance computes the total absolute differences between corresponding components of two color vectors. It emphasizes linear deviations along each dimension. CIE94 Distance (∆E94): Expanding upon ∆E76, the CIE94 metric incorporates perceptual weighting factors that account for variations in lightness, chroma, and hue. These adjustments improve the alignment between numerical color differences and human visual perception [25]. CIEDE2000 Distance (∆E00): Refines ∆E94 by introducing additional perceptual corrections, including chroma and hue interactions. These refinements enhance the accuracy of color difference assessments [8]. For each distance metric, the distances between all sample pixels and the predefined centroids of the respective are computed. The final image classification is achieved based on majority voting across all pixels in an image. This process is independently performed for each distance metric, facilitating multiple assessments of color similarity. C. Threshold determination Selecting an appropriate classification threshold accounts for natural color variability in each wine category. Thresholds were determined by analyzing the distance distributions using statistical metrics, with the median chosen as the most robust approach to minimize outliers. Table III presents the final thresholds for each category. Figure 3 depicts the different CIEDE2000 distance distributions for random sampling using the reference centroids. Each histogram highlights the mean and median values, guiding threshold determination for accurate classification. The selection process ensures that centroids represent each wine category, whereas thresholds classify wines based on perceptually significant differences. VI. RESULTS & DISCUSSION Table IV provides an overview of the precision, recall, F1-score, and accuracy achieved by three distinct classification strategies—Closest-Centroid, Threshold-Based, and SVM-based—across both RGB and LAB color spaces. Within the RGB color space, we explore Euclidean and Manhattan distance measures. In the LAB color space, we examine Manhattan LAB, CIE76, CIE94, and CIEDE2000. The Closest-Centroid method tends to achieve lower performance in the RGB color space. The fact that the RGB space may not accurately represent perceptual color disparities. Specifically, the Euclidean and Manhattan RGB results rarely surpass 0.35 in F1-score, underscoring the inherent limitations of purely distance-based classification within a non-perceptual color space. When transitioning to LAB, this same approach shows substantial improvement, particularly under standard perceptual metrics like CIE76, CIE94, and CIEDE2000. Threshold-Based classification displays mixed performance. In the RGB space, it performs slightly better than ClosestCentroid, indicating that well-defined distance thresholds may capture subtle color variations more effectively. However, the gap between precision and recall can still be considerable. Setting narrower thresholds might amplify detection of certain wine types but risks overlooking others. Likewise, using perceptual metrics in LAB substantially raises overall performance and tends to balance precision and recall, although conservative thresholds sometimes lead to high precision at the expense of recall. The SVM-based approach consistently demonstrates superior performance across nearly all tested scenarios. It achieved F1-scores above 0.90 in RGB and approaching or exceeding 0.95 in LAB with perceptual metrics, adapting more flexibly to complex decision boundaries. This advantage becomes more pronounced when using LAB metrics, where perceptual differences among red, white, and ros´ e are captured more effectively. Notably, under CIEDE2000, the SVM-based classifier attains best precision, recall, and F1-scores. The contributions of this study extend beyond classification accuracy, encompassing aspects of practicality and implementation. An important contribution is the integration of smartphones for both data acquisition and processing. In this context, the smartphone functions not only as an imaging device but also as a computational platform, enabling a portable, cost-effective, and real-time solution for wine classification. As mentioned, this setup allows operation in unconstrained environments, without the need for specialized laboratory equipment. The use of readily available consumergrade smartphones further enhances accessibility, making the approach viable for end-users and small-scale producers. The tests using a Samsung Galaxy A40 (Exynos 7904, 4 GB RAM) Fig. 3. Distributions of distances from each centroid for threshold determination. TABLE III THRESHOLD VALUES FOR WINE CLASSIFICATION USING DIFFERENT DISTANCE METRICS. Wine Type Euclidean (RGB) Manhattan (RGB) CIE76 Manhattan (LAB) CIE94 CIEDE2000 Red 25.4 39.0 9.3 14.9 8.7 7.8 White 68.95 111.0 19.29 27.29 14.89 12.0 Ros´ e60.0 91.0 18.35 28.35 11.33 11.14 TABLE IV PRECISION,RECALL, F1-SCORE,AND ACCURACY FOR EACH COMBINATION OF CLASSIFICATION APPROACH AND METRIC IN RGB AND LAB SPACES. Classification Approach Euclidean RGB Manhattan RGB Manhattan LAB Precision Recall F1-Score Accuracy Precision Recall F1-Score Accuracy Precision Recall F1-Score Accuracy Closest-Centroid 0.308 0.300 0.301 0.300 0.222 0.217 0.179 0.217 0.913 0.905 0.904 0.905 Threshold-Based 0.311 0.306 0.306 0.306 0.356 0.309 0.326 0.309 0.968 0.816 0.883 0.816 SVM-based 0.929 0.931 0.930 0.928 0.871 0.876 0.872 0.868 0.951 0.953 0.952 0.951 CIE76 CIE94 CIE2000 Precision Recall F1-Score Accuracy Precision Recall F1-Score Accuracy Precision Recall F1-Score Accuracy Closest-Centroid 0.915 0.908 0.908 0.908 0.879 0.873 0.872 0.873 0.903 0.896 0.896 0.896 Threshold-Based 0.930 0.790 0.850 0.790 0.987 0.796 0.880 0.796 0.974 0.808 0.880 0.808 SVM-based 0.947 0.951 0.949 0.948 0.976 0.977 0.977 0.977 0.979 0.980 0.979 0.980 demonstrated that the complete pipeline including image capture, ROI extraction, pixel sampling, and classification using an SVM model with the CIEDE2000 metric can be performed in an average of 50 milliseconds per image. These experiments highlight how incorporating perceptual color representations and decision-making methods can notably improve classification accuracy. By using color metrics and machine learning techniques, the approaches tested were able to capture distinctions among red, white, and ros´ e wines. VII. CONCLUSION This study presents an image-based method for coarsegrained wine color classification using LAB color space and the CIEDE2000 metric. Outperforming RGB methods, evaluations on red, white, and ros´ e wines show that using optimized color centroids with CIEDE2000 achieves high accuracy despite lighting and acquisition challenges. The proposed approach enables real-time wine color analysis for consumers, visually impaired users, and potential multimodal integration with additional sensing technologies. Future research may extend the scope of wine categories analyzed and integrate advanced calibration techniques to further attenuate environmental influences. Additionally, expanding the dataset—potentially through crowdsourced image contributions—could enhance the model’s robustness against greater intra-class color variability. ACKNOWLEDGMENT This work was conducted as part of the Watson project, funded by the European Union’s Horizon Europe research and innovation programme, under grant agreement No. 101084265. DISCLAIMER Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Executive Agency (REA). Neither the European Union nor the granting authority can be held responsible for them. REFERENCES [1] Marcel Hensel et al. “New Insights into Wine Color Analysis: A Comparison of Analytical Methods to Sensory Perception for Red and White Varietal Wines”. In: Journal of Agricultural and Food Chemistry 72 (2024), pp. 2008–2017. [2] Di Wu and Da-Wen Sun. “Colour measurements by computer vision for food quality control: A review”. In: Trends in Food Science & Technology 29 (2013), pp. 5–20. [3] Maria Lourdes Gonzalez-Miret et al. “Measuring colour appearance of red wines”. In: Food Quality and Preference 18 (2007), pp. 862–871. [4] Carlos Monteiro de Lima et al. “Digital image-based tracing of geographic origin, winemaker, and grape type for red wine authentication”. In: Food Chemistry 312 (2020), p. 126060. [5] Gaurav Sharma, Wencheng Wu, and Edul N. Dalal. “The CIEDE2000 color-difference formula: Implementation notes, supplementary test data, and mathematical observations”. In: Color Research & Application 30.1 (2005), pp. 21–30. DOI: https://doi.org/10.1002/col. 20070. URL: https://onlinelibrary.wiley.com/doi/abs/10. 1002/col.20070. [6] Dominik Durner and Marcel Hensel. “Is there a need to re-define the methods to evaluate wine color”. In: BIO Web of Conferences 68 (2023), p. 02008. [7] Saeedeh Abasi, Mohammad Amani Tehran, and Mark D Fairchild. “Distance metrics for very large color differences”. In: Color Research & Application 30 (2019), pp. 1–16. [8] Am´ erico Pereira et al. “Efficient CIEDE2000-Based Color Similarity Decision for Computer Vision”. In: IEEE Transactions on Circuits and Systems for Video Technology 30.7 (2020), pp. 2141–2154. DOI: 10.1109/ TCSVT.2019.2914969. [9] Ming Luo, Guihua Cui, and B. Rigg. “The development of the CIE 2000 colour-difference formula: CIEDE2000”. In: Color Research & Application 26 (Oct. 2001), pp. 340–350. DOI: 10.1002/col.1049. [10] Pankaj B Pathare, Umezuruike Linus Opara, and Fahad Al-Julanda Al-Said. “Colour measurement and analysis in fresh and processed foods: A review”. In: Food and bioprocess technology 6 (2013), pp. 36–60. [11] Burc¸e Atac¸ Mogol and Vural G¨ okmen. “Computer vision-based analysis of foods: A non-destructive colour measurement tool to monitor quality and safety”. In: Journal of the science of food and agriculture 94 (May 2014). DOI: 10.1002/jsfa.6500. [12] Am´ erico Pereira, Pedro Carvalho, and Luis CorteReal. “Boosting color similarity decisions using the CIEDE2000 PF Metric”. In: Signal, Image and Video Processing 16.7 (2022), pp. 1877–1884. [13] Kassara Stella, Parker Mango, et al. “Phenolic Compositions of 50 and 30 Year Sequences of Australian Red Wines: The Impact of Wine Age”. In: (2012). [14] Mar´ ıa-Pilar S´ aenz-Navajas et al. “Pigment composition and color parameters of commercial Spanish red wine samples: Linkage to quality perception”. In: European Food Research and Technology 232 (2011), pp. 877– 887. [15] Marcel Hensel et al. “Specification and Simplification of Analytical Methods to Determine Wine Color”. In: Processes 10.12 (2022). ISSN: 2227-9717. DOI: 10 . 3390/pr10122707. URL: https://www.mdpi.com/22279717/10/12/2707. [16] AN Underhill, CD Hirsch, and MD Clark. “Evaluating and mapping grape color using image-based phenotyping”. In: Plant Phenomics (2020). [17] Carlos Herrero-Latorre et al. “Detection and quantification of adulterations in aged wine using RGB digital images combined with multivariate chemometric techniques”. In: Food chemistry: X 3 (2019), p. 100046. [18] Mark D. Fairchild. “Visual and photographic assessment of wine color”. In: Color Research & Application 48.1 (2023), pp. 21–31. DOI: https://doi.org/10.1002/col. 22787. eprint: https://onlinelibrary.wiley.com/doi/pdf/ 10.1002/col.22787. URL: https://onlinelibrary.wiley. com/doi/abs/10.1002/col.22787. [19] BANC Roxana et al. “Phenolic composition and antioxidant activity of red, ros´ e and white wines originating from Romanian grape cultivars”. In: Notulae Botanicae Horti Agrobotanici Cluj-Napoca 48.2 (2020), pp. 716– 734. [20] P. G. Miguel de Oliveira. Wine Glass Dataset. Retrieved from Kaggle. 2022. URL: https : / / www. kaggle . com / datasets/migvanderlei/testedatasetwine. [21] Stefan Aeberhard and M. Forina. Wine. UCI Machine Learning Repository. DOI: https://doi.org/10.24432/C5PC7J. 1992. [22] Paulo Cortez et al. “Modeling wine preferences by data mining from physicochemical properties”. In: Decision Support Systems 47.4 (2009). Smart Business Networks: Concepts and Empirical Evidence, pp. 547–553. ISSN: 0167-9236. DOI: https://doi.org/10.1016/j.dss.2009. 05.016. URL: https://www.sciencedirect.com/science/ article/pii/S0167923609001377. [23] Xintao Wang et al. Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data. 2021. arXiv: 2107.10833 [eess.IV].URL: https://arxiv.org/ abs/2107.10833. [24] Christopher M Bishop. “Pattern recognition and machine learning”. In: Springer google schola 2 (2006), pp. 1122–1128. [25] Roderick McDonald and Kenneth J Smith. “CIE94-a new colour-difference formula”. In: Journal of the Society of Dyers and Colourists 111.12 (1995), pp. 376–379.