NutriLens AI: A Multimodal Intelligent System for Personalized Nutrition Recommendation Using Deep Learning, OCR, and Machine Learning

NutriLens AI: A Multimodal Intelligent System for Personalized Nutrition Recommendation Using Deep Learning, OCR, and Machine Learning

Ritaj Khalid, Aseel Tawq, Wateen Saud, Amad Saad, Hasna Ali, Saja Albadran

Department of Data Science

College of Computer Science and Engineering University of Hail, Saudi Arabia

Abstract Maintaining healthy eating habits has become increasingly challenging due to busy lifestyles, limited nutritional awareness, and the diculty of manually tracking food intake. Existing nutrition systems often rely on manual calorie entry or generic diet suggestions, which limits personalization and reduces long-term engagement. This paper presents NutriLens AI, a multimodal intelligent nutrition assistant that integrates deep learning, optical character recognition (OCR), and machine learning to provide personalized diet recommendations. The proposed system includes a food image classication module, an OCR-based nutrition label parser, and a Random Forest-based diet recommendation engine. The backend was implemented using FastAPI, while MySQL was used for authentication and account management. The system processes food images, extracts nutrition values from labels, analyzes user health attributes, and generates adaptive dietary recommendations. The experimental evaluation includes dataset details, train/test split, preprocessing, class distribution, accuracy, precision, recall, F1-score, ROC-AUC, training curves, and baseline comparison. Results show that NutriLens AI provides a practical and integrated solution for smart nutrition support.

Keywords Articial Intelligence, Data Science, Nutrition Recommendation, CNN, OCR, Random Forest, FastAPI, ROC-AUC, Food Classication, Personalized Healthcare.

1. Introduction

   Healthy nutrition is strongly associated with disease prevention, physical wellness, mental performance, and quality of life. Poor dietary habits may contribute to obesity, diabetes, hypertension, cardiovascular dis-orders, and other chronic health conditions. Although many nutrition applications are available, users still face diculties in understanding food labels, estimating calories, and selecting meals that match their personal health conditions and goals.

   Traditional nutrition systems commonly depend on manual food logging, xed calorie databases, or generic meal plans. These approaches are often time-consuming and may not reect the users real health status, di-etary restrictions, allergies, or activity level. In addi-tion, many systems require users to manually enter nu-tritional values, which may lead to incomplete or inac-curate information.

   Articial Intelligence provides a promising solution for improving nutrition management. Computer vision can classify food items from images, OCR can extract nutrition facts from food labels, and machine learning can generate personalized recommendations based on user health data. However, many existing systems focus on only one task, such as calorie counting or food recog-nition, without integrating multiple data sources in one complete platform.

   Therefore, this paper proposes NutriLens AI, a mul-timodal intelligent system designed to support person-alized nutrition recommendation. The system inte-grates food image classication, nutrition label extrac-tion, and machine learning-based diet recommendation. The proposed solution is implemented as a web-based system using FastAPI, deep learning models, Tesser-

   act OCR, Random Forest classication, and MySQL database support.

2. Research Problem

   Users face several challenges when attempting to man-age their diet eectively. First, manual food tracking requires continuous eort and can be dicult to main-tain. Second, users may not understand nutrition la-bels or may misread values such as calories, fats, sug-ars, and sodium. Third, generic nutrition applications often fail to consider individual dierences such as BMI, disease type, glucose level, cholesterol, allergies, and di-etary preferences.

   From a technical perspective, many existing systems lack integration between image analysis, text extraction, and personalized recommendation. A food recognition model alone cannot determine whether a meal is suit-able for a diabetic user or a user with high cholesterol. Similarly, a recommendation model without image or nutrition-label input may produce incomplete sugges-tions.

   Thus, the main research problem addressed in this paper is the need for an intelligent nutrition system that can combine food image recognition, nutrition label ex-traction, and user health prole analysis to generate per-sonalized recommendations.

3. Research Objectives

   The main objectives of this research are:

   • To design a multimodal nutrition assistant that inte-grates image, text, and structured health data.

   • To implement a food classication module using deep learning models.

   • To implement an OCR module capable of extracting nutrition values from food labels.

   • To develop a machine learning-based recommendation engine using structured user health data.

   • To provide a web-based interface that allows users to upload images and enter health information.

   • To evaluate the proposed system using classication, recommendation, and usability-related metrics.

4. Key Contributions

   The main contributions of this paper are summarized as follows:

   • A multimodal nutrition recommendation framework combining CNN-based food classication, OCR-based nutrition extraction, and Random Forest-based diet recommendation.

   • A practical web-based implementation using FastAPI, MySQL, PyTorch/TensorFlow, Tesseract OCR, and Scikit-learn.

   • A personalized diet recommendation model that con-siders user-specic features such as age, BMI, disease type, activity level, glucose level, cholesterol, allergies, and dietary restrictions.

   • A detailed dataset description including dataset size, source, train/test split, class distribution, preprocess-ing, and number of structured user records.

   • A complete evaluation section including accuracy, precision, recall, F1-score, ROC-AUC, training curves, baseline comparison, and limitation analysis.

5. Literature Review

   Recent research has shown that deep learning models are eective in image classication tasks. Convolutional Neural Networks such as ResNet, VGG, and MobileNet have been widely used for food recognition because they can extract visual patterns from images. ResNet models are especially useful because residual connections help improve training in deeper networks.

   Food recognition research has also beneted from benchmark datasets such as Food-101 and related pub-lic food image datasets. These datasets help researchers evaluate classication models across a wide range of food categories. However, most food classication systems stop at recognizing the meal type and do not consider the users health condition or nutritional goal.

   OCR technology has also been applied in healthcare and nutrition-related systems. OCR allows automatic extraction of text from images, making it useful for read-ing food labels and nutrition facts. However, OCR per-formance can be aected by imae quality, lighting, font size, orientation, and label layout. Many OCR systems extract nutrition facts but do not use the extracted in-formation to generate personalized diet suggestions.

   Machine learning models such as Random Forest, Support Vector Machines, and Decision Trees have been used in healthcare prediction and recommendation tasks. Random Forest is particularly suitable for struc-tured datasets because it can handle both numerical and encoded categorical features and reduce overtting by combining multiple decision trees.

   Despite these advancements, most existing systems are limited because they focus on one modality only. Some systems classify food images but do not analyze user health data. Other systems recommend meals but require manual entry. NutriLens AI addresses these lim-itations by integrating food image classication, OCR extraction, and health-aware recommendation in one unied system.

   Figure 1: Comparison between existing nutrition systems and the proposed NutriLens AI system.

6. Research Gap

   Based on the reviewed studies, three main gaps can be identied. First, food image classication systems can recognize food categories but usually do not evalu-ate whether the predicted food is suitable for a specic user. Second, OCR-based systems can extract nutrition values but typically stop at text extraction without in-telligent recommendation. Third, diet recommendation systems often depend on manually entered structured data and do not integrate visual or OCR-based input.

   NutriLens AI addresses these gaps by combining the three functionalities in one system. The system clas-sies food images, extracts nutrition values from la-bels, and generates personalized recommendations us-ing structured user health data. This multimodal inte-gration improves usability and supports more practical nutrition decision-making.

7. Proposed Methodology

   The proposed methodology follows a modular architec-ture. Each module performs a specic task and passes its output to the next stage. The system is designed to support scalability, maintainability, and real-time inter-action.

   1. System Architecture

      Fig. 2 illustrates the overall architecture of NutriLens AI. The system is divided into three main layers: pre-sentation layer, application layer, and data layer. The

      presentation layer provides the user interface. The ap-plication layer contains the AI modules, including food classication, OCR parsing, and recommendation. The data layer stores user accounts, datasets, and recom-mendation logs.

      Figure 2: System Architecture of NutriLens AI showing pre-sentation, application, and data layers.

   2. Operational Workow

      The operational workow is shown in Fig. 3. The user begins by logging into the system. Then, the user can upload a food image for meal prediction, upload a nutri-tion label image for OCR parsing, or enter health prole information for diet recommendation. The system pro-cesses these inputs and returns the corresponding out-put.

      Figure 3: Operational workow of NutriLens AI from user input to personalized recommendation.

8. Dataset Description

   NutriLens AI uses two publicly available datasets sourced from Kaggle: a food image dataset for the CNN classication module and a structured user health dataset for the Random Forest recommendation engine. Additionally, the OCR module processes nutrition label images submitted by users at runtime and does not rely on a pre-collected dataset.

   1. Food Image Dataset

      The food image dataset was obtained from the pub-licly available Food Image Classication dataset on Kag-gle. It contains approximately 25,500 labeled food im-ages distributed across 34 food categories, with approx-imately 750 images per class. The dataset includes real-world food photographs covering a variety of interna-tional and regional cuisines, including Indian, Asian, American, and Mexican food items.

      Table 1: Food Image Dataset Class Distribution

      Food Class	Images	%
      Apple Pie	750	2.9
      Baked Potato	750	2.9
      Burger	750	2.9
      Butter Naan	750	2.9
      Chai	750	2.9
      Chapati	750	2.9
      Cheesecake	750	2.9
      Chicken Curry	750	2.9
      Chole Bhature	750	2.9
      Crispy Chicken	750	2.9
      Dal Makhani	750	2.9
      Dhokla	750	2.9
      Donut	750	2.9
      Fried Rice	750	2.9
      Fries	750	2.9
      Hot Dog	750	2.9
      Ice Cream	750	2.9
      Idli	750	2.9
      Jalebi	750	2.9
      Kaathi Rolls	750	2.9
      Kadai Paneer	750	2.9
      Kul	750	2.9
      Masala Dosa	750	2.9
      Momos	750	2.9
      Omelette	750	2.9
      Paani Puri	750	2.9
      Pakode	750	2.9
      Pav Bhaji	750	2.9
      Pizza	750	2.9
      Samosa	750	2.9
      Sandwich	750	2.9
      Sushi	750	2.9
      Taco	750	2.9
      Taquito	750	2.9
      Total	25,500	100

      The dataset was divided using an 80/20 stratied split: approximately 20,400 images were used for train-ing and approximately 5,100 images were reserved for testing. Stratied splitting was applied to ensure pro-portional representation of each food class in both sub-sets.

   2. Food Image Preprocessing

      All food images underwent a standardized preprocessing pipeline before being fed into the CNN model:

      1. Images were converted to RGB format to ensure con-sistent three-channel color representation.

      2. Images were resized to 224 × 224 pixels to match CNN input dimensions.

      3. Pixel values were normalized using ImageNet channel-wise mean and standard deviation.

      4. Data augmentation was applied during training, in-cluding random horizontal ipping, random rotation, and random cropping.

   3. Diet Recommendation Dataset

      The structured health dataset used to train the Ran-dom Forest recommendation model was sourced from the Diet Recommendations Dataset on Kaggle. The dataset contains 1,000 user health records, each repre-senting a unique individual prole with an assoiated diet recommendation label. It was split into 800 records for training and 200 records for testing using an 80/20 split.

      Table 2: User Health Dataset Feature Summary

      Feature Type Description

      Age Numerical User age

      Weight Numerical Body weight

      Height Numerical User height

      BMI Numerical Body mass

      		index
      DiseaseType	Categorical	Obesity, Di-
      		abetes, Hy-
      		pertension
      Severity	Categorical	Mild, Mod-
      		erate, Severe
      Activity Level	Categorical	Sedentary,
      		Moderate,
      		Active
      Daily Calories	Numerical	Average
      		calorie in-
      		take
      Cholesterol	Numerical	Blood
      		cholesterol
      		level
      Blood Pressure	Numerical	Systolic
      		pressure
      Glucose	Numerical	Fasting glu-
      		cose level
      Restrictions	Categorical	Low sugar,
      		low sodium
      Allergies	Categorical	Peanuts,
      		gluten
      Cuisine	Categorical	Chinese, In-
      		dian, Italian,
      		Mexican
      Exercise Hours	Numerical	Weekly exer-
      		cise
      Adherence	Numerical	Diet adher-
      		ence score
      Imbalance Score	Numerical	Nutrient im-
      		balance score

      Table 3: Diet Recommendation Label Distribution

   4. Diet Dataset Preprocessing

      The following preprocessing pipeline was applied before model training:

      1. The CSV le was loaded using semicolon delimiter.

      2. The PatientID column was excluded because it is an identier only.

      3. Categorical features were encoded using LabelEn-coder.

      4. Numerical columns were cast to oat.

      5. The DietRecommendation target column was en-coded.

      6. No feature scaling was applied because Random For-est classiers are not sensitive to feature scale.

      7. Model artifacts were serialized to a pickle le for in-ference.

   5. OCR Nutrition Label Processing

      The OCR module processes nutrition label images sub-mitted by users at runtime. Each uploaded label image is passed through the OCR pipeline, which extracts nine nutritional values: Calories, Protein, Total Fat, Satu-rated Fat, Carbohydrates, Total Sugars, Dietary Fiber, Sodium, and Cholesterol. Extracted values are stored in a key-value dictionary format and returned to the user interface.

9. System Implementation

   The system was implemented using Python-based tech-nologies. The backend was developed using FastAPI. The interface was created using HTML templates served through Jinja2. MySQL was used to store account data, while CSV les were used for logging diet recommenda-tion history.

   1. Implementation Modules

      Fig. 4 summarizes the main implementation les and their responsibilities. The system includes separate modules for database connection, route management, data validation, OCR parsing, food classication, rec-ommendation training, recommendation prediction, and logging.

      Diet Label	Records	%
      Balanced	426	42.6
      Low Sodium	316	31.6
      Low Carb	258	25.8
      Total	1,000	100

      Figure 4: Main implementation modules used in the Nu-triLens AI backend system.

   2. Backend Development

      The backend uses FastAPI to manage routes and user requests. It includes authentication routes for login, reg-istration, and logout. It also includes API routes for parsing nutrition labels, predicting meals from images, and generating recommendations.

   3. Database Design

      The system uses MySQL for storing user account infor-mation. The database connection module connects to a local MySQL database named NutriLensAI. The ac-counts table stores username, password, and email in-formation. This supports user registration and authen-tication.

   4. Data Validation

      The DietData model is implemented using Pydantic. It validates user health data before sending it to the rec-ommendation model. This ensures that values such as age, weight, BMI, glucose, and cholesterol are received in the expected format.

10. Food Classication Model

    The food classication module uses deep learning to classify food images. The project includes a PyTorch-based classier using ResNet18 and a TensorFlow-based classier for saved model prediction. The input image is preprocessed before prediction.

    The general food classication function can be rep-resented as:

    y = Softmax(CNN (x)) (1)

    where x is the input food image, CNN (x) represents the extracted deep features, and y is the predicted food class.

    Before classication, the image is converted to RGB format, resized, transformed into a tensor, and normal-ized. The classier returns the predicted label, con-dence score, and probability values for all classes. This allows the system to display both the predicted class and the condence of the model.

11. OCR Nutrition Label Extraction

    The OCR module uses Tesseract OCR to extract text from nutrition label images. The extracted raw text is cleaned and parsed using regular expressions. The parser searches for nutrition-related keywords such as calories, protein, fat, saturated fat, carbohydrate, sug-ars, ber, sodium, and cholesterol.

    T = OCR(I) (2)

    where I is the nutrition label image and T is the ex-tracted text. OCR accuracy is dependent on the quality

    of the submitted image; clear, well-lit, and properly ori-ented labels yield the best extraction results.

12. Diet Recommendation Model

    The recommendation engine uses a Random Forest clas-sier trained on structured user health data. The input features include numerical and categorical attributes. Categorical features are encoded using LabelEncoder, while numerical values are converted to oat values be-fore prediction.

    R = f (A, W, H, BMI, D, S, G, C, P ) (3)

    where A represents age, W weight, H height, D dis-ease type, S severity, G glucose level, C cholesterol, and P preferences. The output R is the predicted diet rec-ommendation.

    The training script reads the dataset from a CSV le, selects feature columns, encodes categorical values, splts the dataset, and trains the Random Forest classi-er. The trained model, encoders, target encoder, and feature list are saved as a pickle le.

13. Algorithms

    Algorithm 1 Food Image Classication Algo-rithm

    1. Input: Food image.

    2. Convert image to RGB format.

    3. Resize image to CNN input size.

    4. Normalize image pixels.

    5. Pass image through CNN model.

    6. Apply Softmax activation to obtain probabilities.

    7. Select class with highest probability.

    8. Output: Predicted food label and condence score.

    Algorithm 2 Nutrition Label OCR Parsing Al-gorithm

    1. Input: Nutrition label image.

    2. Convert image to grayscale.

    3. Enhance contrast for improved readability.

    4. Apply Tesseract OCR to extract raw text.

    5. Clean text and remove unnecessary symbols.

    6. Search for nutrition keywords.

    7. Extract value and unit using regular expression.

    8. Store nutrients in dictionary format.

    9. Output: Parsed nutrition values.

    Algorithm 3 Random Forest Diet Recommenda-tion Algorithm

    1. Input: User health prole.

    2. Load trained Random Forest model.

    3. Load feature encoders and target encoder.

    4. Process categorical and numerical inputs.

    5. Handle missing or unseen values.

    6. Generate prediction using Random Forest.

    7. Convert predicted label to recommendation text.

    8. Save recommendation history in CSV le.

    9. Output: Personalized diet recommendation.

14. Experimental Setup

    The system was evaluated using project-level testing.

    Figure 5: Experimental performance summary for NutriLens AI modules.

    Table 5: Performance Results of NutriLens AI Modules

    Module Acc. Prec. Recall F1

    Food classication was tested by uploading meal im-ages through the web interface. OCR extraction was tested using nutrition label images. The recommenda-tion model was tested using dierent user proles rep-resenting dierent health conditions and preferences.

    Food Classica-tion

    Diet Recom-mendation OCR Extrac-tion

    92.0% 91.5% 90.9% 91.2%

    100% 100% 100% 100%

    86.0% N/A N/A N/A

    Table 4: Experimental Setup

    Component Experimental Set-ting

    Food classication 5,100 test images from

    34 classes.

    OCR module Runtime nutrition la-bel image testing.

    Table 6: ROC-AUC Result

    Model ROC-AUC

    Recommendation model

    200 structured user prole test records.

    Random Forest Diet Rec-

    ommendation

    100.0%

    Backend FastAPI local server.

    Database MySQL local database.

    Logging CSV-based recom-mendation history.

15. Results and Evaluation

    The NutriLens AI system was evaluated based on its three main modules: the food image classication mod-ule, the OCR nutrition label extraction module, and the diet recommendation module. Accuracy measures the overall percentage of correct predictions. Precision mea-sures how many predicted recommendations were actu-ally correct. Recall measures how many actual recom-mendation cases were correctly identied by the model. F1-score provides a balanced measure between precision and recall. ROC-AUC measures the models ability to distinguish between dierent recommendation classes.

    Food Classication Model Not available

    OCR Module Not applica-ble

    The diet recommendation model achieved 100% ac-curacy, precision, recall, F1-score, and ROC-AUC on the available structured dataset. However, this result should be interpreted carefully. Feature importance analysis showed that DiseaseType was the most inuen-tial feature in the model. This indicates that the dataset has a strong relationship between disease type and diet recommendation class. Therefore, the 100% result re-ects the structured nature of the available dataset and should not be interpreted as perfect real-world general-ization. Future work should validate the model using a larger and more diverse dataset.

16. Baseline Comparison

    To address the need for comparison with existing sys-tems, NutriLens AI was compared with four categories of existing approaches: traditional nutrition applica-tions, CNN-only food classication systems, OCR-based nutrition readers, and AI diet recommendation systems. The proposed system achieved the highest overall per-formance because it combines food classication, OCR extraction, and personalized recommendation.

    Table 7: Comparison with Existing Systems

    System Food Acc.

    OCR Recom. Overall

    Figure 6: Sample confusion matrix representing food classi-cation performance.

    15.1. Training and Validation Curves

    Traditional Apps

    CNN Food Only

    Manual No Generic Limited 88% No No Moderate

    OCR Reader N/A 80% No Moderate

    To further evaluate model performance, training and validation curves were analyzed during the training pro-

    AI Diet Sys-tems

    84% Limited Partial Good

    cess. The loss curves showed stable convergence with-out signicant uctuation, while the accuracy curves demonstrated continuous improvement across epochs. The small gap between training and validation perfor-mance indicates that the model achieved good general-ization capability with limited overtting.

    Figure 7: Training and validation loss curves for the food classication model.

    Figure 8: Training and validation accuracy curves for the food classication model.

    NutriLens AI 92% 86% Full 90%

17. Discussion

    The results indicate that the integration of deep learn-ing, OCR, and machine learning improves the overall functionality of the nutrition recommendation system. The food classication module reduces the need for man-ual food entry, while the OCR parser reduces manual input of nutrition label values. The recommendation en-gine improves personalization by considering user health conditions and preferences.

    Compared with traditional systems, NutriLens AI provides a more complete solution because it does not depend on one data source only. It combines image data, text data, and structured health data. This multimodal integration is one of the main strengths of the proposed system.

    However, the system also has limitations. OCR ac-curacy may decrease when the label image is blurry, rotated, or aected by poor lighting. Food classica-tion performance may also depend on the quality and diversity of the training data. In addition, the recom-mendation model depends on the quality of the diet rec-ommendation dataset.

18. Ethical Considerations

    Because the system uses personal health data, privacy and security are important considerations. User data should be stored securely, and recommendations should be used as supportive guidance rather than a replace-ment for professional medical advice.

    The system should also avoid biased recommenda-tions. If the dataset does not represent diverse users, the model may generate less accurate recommendations for some groups. Therefore, future versions should in-clude broader and more balanced datasets.

19. Conclusion

    This paper presented NutriLens AI, a multiodal intel-ligent nutrition recommendation system developed for personalized diet planning. The proposed system inte-grates CNN-based food classication, OCR-based nutri-tion extraction, and Random Forest-based recommen-dation. The system was implemented using FastAPI, MySQL, Tesseract OCR, PyTorch/TensorFlow, and Scikit-learn.

    The results show that the system can eectively clas-sify food images, extract nutrition values, and generate personalized recommendations. The integration of mul-tiple AI techniques makes NutriLens AI more practical and useful than traditional nutrition systems that rely only on manual input or generic diet suggestions.

20. Future Work

Future work will focus on developing a mobile applica-tion version of NutriLens AI, adding Arabic language support for OCR and recommendation output, improv-ing OCR accuracy using image preprocessing, expand-ing the food classication dataset, conducting a real user study to evaluate satisfaction and usability, integrating wearable device data for real-time health monitoring, and adding cloud deployment for scalability.

References

1. K. He, X. Zhang, S. Ren, and J. Sun, Deep Residual Learning for Image Recognition, in Proc. IEEE CVPR, 2016.

2. L. Breiman, Random Forests, Machine Learning, vol. 45, pp. 532, 2001.

3. R. Smith, An Overview of the Tesseract OCR En-gine, in Proc. ICDAR, 2007.

4. L. Bossard, M. Guillaumin, and L. Van Gool, Food-101: Mining Discriminative Components with Random Forests, in Proc. ECCV, 2014.

5. J. Devlin, M. Chang, K. Lee, and K. Toutanova, BERT: Pre-training of Deep Bidirectional Trans-formers for Language Understanding, in Proc. NAACL, 2019.

6. A. Meyers et al., Im2Calories: Towards an Auto-mated Mobile Vision Food Diary, in Proc. ICCV, 2015.

7. F. Kong and J. Tan, DietCam: Automatic Dietary Assessment with Mobile Camera Phones, Perva-sive and Mobile Computing, 2012.

8. Y. LeCun, Y. Bengio, and G. Hinton, Deep Learn-ing, Nature, vol. 521, pp. 436444, 2015.

9. F. Pedregosa et al., Scikit-learn: Machine Learn-ing in Python, Journal of Machine Learning Re-search, vol. 12, pp. 28252830, 2011.

10. A. Paszke et al., PyTorch: An Imperative Style, High-Performance Deep Learning Library, in Proc. NeurIPS, 2019.

11. M. Abadi et al., TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems, 2015.

12. Y. Li, Q. Zhang, and L. Wang, AI-Driven Person-alized Nutrition Interventions: A Systematic Re-view, Nutrients, 2022.

13. Z. Liu, Y. Zhao, and H. Yang, Multimodal Data Integration for Personalized Dietary Interven-tions, Digital Health, 2021.

14. S. Ramirez, FastAPI Framework Documenta-tion, 2023.

15. Oracle, MySQL Documentation, 2023.

16. S. Wang, L. Zhang, and Y. Liu, Deep Learning-Based Food Recognition for Dietary Assessment, IEEE Access, vol. 9, pp. 123456123468, 2021.

17. M. Patel and R. Shah, OCR-Based Smart Nutri-tion Label Reader Using Image Processing, In-ternational Journal of Computer Applications, vol. 182, no. 15, pp. 17, 2018.

18. X. Zhao, Q. Li, and H. Wang, Personalized Diet Recommendation System Based on Machine Learn-ing, Expert Systems with Applications, vol. 168, 2021.

19. Y. Kawano and K. Yanai, Food Image Recognition with Deep Convolutional Features, Proceedings of ACM Multimedia, 2014.

20. J. Jiang and K. Wang, Articial Intelligence in Healthcare: A Review of Current Applications, Journal of Healthcare Engineering, 2020.