Gluten identification from food images using advanced deep learning and transfer learning methods.

Gluten identification from food images using advanced deep learning and transfer learning methods - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice J Food Sci Technol . 2024 Dec 31;62(6):1164–1172. doi: 10.1007/s13197-024-06158-y Gluten identification from food images using advanced deep learning and transfer learning methods Mayura Tapkire 1, 2 , Vanishri Arun 3, ✉ , M S Lavanya 3 , R Shashidhar 4 Author information Article notes Copyright and License information 1 Department of Information Science and Engineering, The National Institute of Engineering, Mysore, India 2 Research Scholar at Department of Information Science and Engineering, JSS Science and Technology University, Mysore, Karnataka India 3 Department of Information Science and Engineering, JSS Science and Technology University, Mysore, Karnataka India 4 Department of Electronics and Communication Engineering, JSS Science and Technology University, Mysore, Karnataka India ✉ Corresponding author. Revised 2024 Nov 27; Accepted 2024 Dec 8; Issue date 2025 Jun. © Association of Food Scientists & Technologists (India) 2024. Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law. PMC Copyright notice PMCID: PMC12078738  PMID: 40386197 Abstract Food image recognition has become an essential application in computer vision, with significant implications for dietary management, particularly for individuals with specific dietary restrictions. This paper shows a novel approach for gluten image classification, designed to assist individuals with celiac disease in identifying gluten-containing foods. Our proposed model leverages a Convolutional Neural Network (CNN) architecture, specifically utilizing a EfficientNet pretrained model, to accurately identify and classify food images. In the proposed model We utilized a curated dataset from the Food101 dataset, selecting 20,000 images focused on common food recipes. We used the EfficientNet pretrained model, achieving a training accuracy of 99.02% and a validation accuracy of 98.38%. The model was further evaluated on 2000 test images, obtaining an accuracy of 99%. The data was meticulously labelled to ensure high-quality training as well as testing processes. Our results demonstrate the model’s effectiveness in gluten classification, highlighting its potential utility for celiac patients. This work contributes to the growing field of food image recognition and offers a valuable tool for dietary management in celiac patients. Keywords: Convolutional neural networks(CNN), Transfer learning, Gluten image classification, EfficientNet Introduction Food image recognition is crucial for health, nutrition, and food safety. Accurate classification aids dietary management, allergy detection, and nutritional assessment. It is especially significant for those with dietary restrictions, such as celiac disease patients, who must avoid gluten to prevent adverse health effects (Stoleru et al. 2022 ). Identifying gluten-containing foods can be challenging, highlighting the need for reliable food classification systems (Tapkire and Arun 2022 ). Advances in deep learning and CNNs offer high accuracy in image classification, making these tasks more efficient. Celiac disease is on the rise globally and requires technological and medical interventions (Zammit et al. 2023 ). Using convolutional neural networks techniques to predict celiac disease at an early stage is particularly relevant, as it allows for timely intervention and the initiation of remedies upon diagnosis, thereby preventing potential long-term health damage to individuals (Arun et al. 2024 ). As a branch of artificial intelligence, deep learning analyses intricate patterns and structures in data by using deep computational models with several levels of abstraction (Tapkire and Arun 2023 ; LeCun et al. 2015 ). The advent of deep convolutional networks has led to significant breakthroughs in image, video, speech, and audio processing (Hinton et al. 2012 ; Tompson et al. 2014 ; Krizhevsky et al. 2017 ). In this work, we make use of recent developments in deep learning to construct a system that can analyse input images and identify whether or not food products contain gluten. Our research aims to provide celiac patients with crucial information about the food they consume, helping to prevent possible life-threatening allergic reactions. This paper suggests an innovative approach to gluten image classification using a pretrained EfficientNet CNN model. EfficientNet has demonstrated superior performance in various image recognition tasks due to its scalable and efficient architecture. Our model is designed to identify and classify food images, distinguishing between gluten and non-gluten items. We trained and validated our model using the Food101 dataset, achieving a training accuracy of 99.02% and obtained a validation accuracy of 98.38% across 20 classes. The methodology section of this work addresses model design, training and evaluation metrics, and data pre-processing. We also discuss the potential enhancements for scalability and accuracy, such as developing deeper neural networks, expanding the dataset with more images per class, and fine-tuning the model’s hyper parameters. This work attempts to foster the growing field of food image recognition and provide a valuable tool for dietary management in celiac patients. Literature review The literature on food image classification highlights the diverse methodologies and datasets used to enhance model accuracy. In one study, the Korean food image dataset was analysed using image augmentation and deep convolutional neural networks (DCNN). The K-foodNet model, when compared to AlexNet, GoogLeNet, VGG, and ResNet, achieved a notable accuracy of 91.3% with a recognition time of 0.4 ms, outperforming other networks. However, the study faced several limitations: a restricted number of high-quality, specific food images, the inclusion of only 23 Korean food categories, and the absence of a suitable public image dataset for Korean food recognition. High-quality images, captured with good lighting and angles, are crucial for enhancing the performance of deep convolutional neural networks (DCNNs). Improving image quality and presentation is essential for achieving better results with deep models (Park et al. 2019 ). The recognition of meal images is crucial for extracting health-related information. Food recognition presents challenges due to the diverse shapes and appearances of dishes, particularly with Indonesian cuisine, where variations exist even within the same food type based on origin. This study introduces a recognition method for Indonesian food using SIFT, K-Dimensional Tree, and Backpropagation Neural Networks (BPNN). BPNN demonstrated higher recognition accuracy at 51% compared to KD Tree’s 44%, although the accuracy remains low due to significant image variation. To improve accuracy, especially for Indonesian foods, larger datasets are necessary (Giovany et al. 2017 ). Diverse food image datasets have made substantial use of various convolutional neural networks (CNNs). This paper addresses food image classification using convolutional neural networks, implemented both from scratch and with transfer learning, achieving an accuracy of 61.4% and a top-5 accuracy of 85.2%. Future work will focus on optimizing hyperparameters and deciding which layers to freeze or train during transfer learning. Due to limited computing resources and time, decisions were based on model convergence and training-validation metrics, although a comprehensive hyperparameter search would be ideal (Islam et al. 2018 ). Similarly, the Traditional Bengali Food Image (TBFI) dataset was analysed using both scratch CNN models and fine-tuned VGG16, resulting in a high accuracy of 98%.These results demonstrate the model’s effectiveness for traditional Bengali food classification and its potential for other food scenarios. Future work could focus on classifying food ingredients for dietary purposes, with data augmentation and enhanced regularization methods further improving results. Additionally, plans include applying the novel dataset to other pretrained models for broader applications (Uddin et al. 2021 ). Furthermore, an ensemble CNN model applied to a customized food recognition dataset comprising 29 varieties yielded a robust accuracy of 95.55%. To improve the model’s performance in food recognition, future work should explore advanced techniques and combine diverse datasets for comprehensive evaluation. Testing on larger datasets will assess its real-world applicability. For practical use in nutrition tools, deploying the model will enhance public health monitoring. Boosting interpretability with techniques like saliency maps will build trust. Collaborating with food scientists and utilizing micro-structure analysis can refine recognition systems, advancing dietary assessments and public health (Abiyev and Adepoju 2024 ). The UEC-Food100 dataset has been extensively studied with various models, including fine-tuned Deep CNN CaffeeNet, which achieved an accuracy of 94.85%, and DenseNet-121, which reached 87.1%.In this paper, the authors explored pretraining and fine-tuning a DCNN on a small-scale food dataset with approximately 100 training images per category. Furthermore, applying the optimized DCNN to Twitter photo data improved food photo mining in both quantity and accuracy. The DCNN’s suitability for large-scale data is evidenced by its rapid classification speed of 0.03 s per image using a GPU. Future work involves adapting the framework for mobile devices by reducing DCNN parameters, totaling about 60 million. Additionally, expanding Twitter food photo mining to analyze food distribution and preferences through geo-spatial and temporal aspects presents a valuable research gap (Yanai and Kawano 2015 ). This paper presents a comparative analysis of deep learning methods for food image recognition, using the UEC Food-100 dataset. Models like ResNet-18, Inception-V3, ResNet-50, DenseNet-121, Wide ResNet-50, and ResNeXt-50 were tested with image sizes of 320 × 320 and 299 × 299. Due to the limited dataset size, a transfer learning approach with ImageNet weights was employed. ResNeXt-50 achieved the highest accuracy at 87.7%. The study suggests that using an EfficientNet pretrained model could further improve accuracy (Memiş et al. 2020 ). Additionally, the MLC-41 dataset was analysed with AlexNet, CaffeeNet, and ResNet-50, where ResNet-50 outperformed others with an accuracy of 87.78%. Designed for small to medium datasets, it can be adapted for large-scale data (Pan et al. 2017 ). Another study using the Pittsburgh Fast-Food Image Dataset with SIFT features attained an accuracy of 82%. This paper introduces a food image classification technique using SIFT for local texture detection, LBP for encoding, and shape context for global structure representation, validated on the Pittsburgh Fast-Food Image dataset, with future work aimed at developing view-invariant texture features to address viewpoint variability (Zong et al. 2010 ). Several research efforts have focused on the Food101 dataset, a popular choice for food image classification. Techniques such as Random Forest Component Mining yielded a lower accuracy of 50.76%, while transfer learning approaches with models like Inception-ResNet and Inception V3 significantly improved performance, achieving Top-1 accuracy of 72.55% and Top-5 accuracy of 91.31%. The method also performs competitively on the MIT-Indoor dataset. The research identifies a gap in further improving performance, particularly in comparison with CNNs, and enhancing part-based classification techniques (Bossard et al. 2014 ; Yu et al. 2016 ). EfficientNetB0 model applied to the same dataset reached an accuracy of 80. The study highlights the challenge of recognizing a broad range of foods in a single model. A research gap exists in exploring the integration of multimodal datasets, including text, image, and audio, to further enhance food recognition capabilities (VijayaKumari et al. 2022 ), and other CNN-based methods showed Top-1 accuracy of 61.4% and Top-5 accuracy of 85.2% (Jiang 2019 ). This study uses deep learning models for reliable food recognition, combining image analysis with semantically relevant phrases from extensive online literature. Tested on a dataset of 100 classes with about 1000 images each, the approach achieved an 85% top-1 classification rate. It extends the benchmark Food-101 dataset to include sub-continental foods, demonstrating effectiveness across both datasets. The solution has been implemented as a mobile app for healthcare. A research gap remains in further developing the Android app by adding capabilities to provide a comprehensive guide for everyday meals (Bahadur et al. 2022 ). A pre-trained Inception v3 CNN model via transfer learning is used to enhance a custom CNN framework, applies data augmentation to improve model robustness and prevent overfitting, and deploys web crawlers for retrieving food-related information like origin and nutritional details. Built with Python Scrapy and automated with Selenium, the system achieves 97% accuracy across 20 classes. A research gap exists in improving scalability by developing a more advanced neural network, gathering additional images per class, fine-tuning hyperparameters, and implementing advanced web crawling techniques to reduce latency and expand search capabilities (Chaitanya et al. 2023 ). A framework for food image classification and recipe extraction using the MResNet-50, a lightweight CNN model that effectively identifies intra-class variability and inter-class similarity is introduced. A novel approach for ingredient identification through NLP and Ontology is proposed, with extensive experimental validation. Its lightweight nature facilitates integration with wearable devices, supporting nutrition tracking and telehealth services. Future exploration could focus on enhancing real-time performance and broadening mobile application possibilities. These studies collectively demonstrate the effectiveness of different deep learning models and methodologies in food image classification, with varying levels of success depending on the dataset and approach used. Methodology Food101 dataset Food101 dataset includes 101 classes of real-world food images, all of which were manually captured and labelled. It includes a total of 101,000 images, with precisely 1,000 images per class. The dataset is divided into a training set with 750 images and a testing set with 250 images for each class, as organized by its creators. Due to its large size and balanced class distribution, the Food101 dataset is often utilized as a primary source for various applications (Ao and Ling 2015 ; Ponce and León 2020 ). Twenty random classes out of 101 were considered, focusing exclusively on commonly used food recipes. Out of 20 classes 10 classes are of Gluten containing recipes like samosa, spring rolls, garlic bread, pizza etc., whereas 10 classes of non-gluten recipes like beet-salad, fried rice, frozen-yogurt, omelette. Each class contains 1000 images. We have divided the images from each class in 87%-13% for training and testing purpose whereas training images were further divided into 62%-25% for training and validation purpose. Model was successfully deployed by achieving an impressive training accuracy 99.02% and validation accuracy 98.38% across 20 classes. To further evaluated the model on test dataset and obtained an accuracy of 99% on 2000 images. The data was meticulously labelled to ensure high-quality training and testing processes. The methodology is illustrated in Fig. 1 , and examples of random gluten and nongluten food images from selected classes are shown in Fig. 2 . A random function was employed to select and display 9 images from both the training and validation sets of the Food-101 dataset. Training of the CNN model was conducted on Google Colab’s Tesla T4 GPU, chosen for its significant speed advantage over conventional CPUs. The proposed model was developed and evaluated within the Google Colab notebook environment, selected for its user-friendly interface and extensive support for machine learning tools. Fig. 1. Open in a new tab Methodology Fig. 2. Open in a new tab Random Images from few classes The gluten classification model using the Food101 dataset involves a structured approach starting with image preprocessing and data augmentation using the PyTorch framework. The training and validation processes were conducted using the PyTorch framework. Before being fed into the model, input images were normalized and resized to 224 × 224 pixels. Data loaders were configured to handle batches of 32 images for both training and validation phases. The training phase was carried out for 20 epochs. During each epoch, the model was trained and validated on the training and validation dataset respectively. The loss function used for optimization was “Cross Entropy Loss”, suitable for multi-class classification tasks. The Adam optimizer, set to learning rate of 0.001 and weight decay to prevent overfitting, was employed to minimize the loss. In order to promote more efficient model convergence, a learning rate scheduler called “StepLR” was also used, which lowers the learning rate by a factor of 0.1 every seven epochs. A number of performance parameters were monitored during the training process, including F1 score, accuracy, precision, recall, and loss while training and validation. The model’s overall correctness is measured by accuracy, and its ability in accurately classifying the Gluten and Non-Gluten classes is demonstrated by precision, recall, and F1 score. The F1 score is calculated as the harmonic mean of precision and recall scores that shows a fair evaluation of the model’s performance. Recall is the percentage of true positive predictions among all actual positives, whereas precision shows the percentage of true positive predictions over all positive predictions. Transfer learning with efficientnet model Convolutional and pooling layers make up a CNN (Park et al. 2019 ), providing a strong architecture for image recognition and classification (Abiyev and Adepoju 2024 ). As the network’s output classifier, the last layer is fully connected. This layer uses an activity regularizer to penalize the model according to the magnitude of activation, hence reducing overfitting. As a result, this layer successfully recovers and acquires features from the original input image that has been through several filter processing steps. The model uses these learned features to do classification. The EfficientNet-B0 model is shown in Fig. 3 . Due to its state-of-the-art performance in various image recognition challenges, the EfficientNet-B0 architecture was selected for this classification task (ElMOUFIDI and Amoun 2021 ). The pre-trained EfficientNet-B0 model was fine-tuned for binary classification by modifying the final fully connected layer to output two classes: “Gluten” and “NonGluten”. Fine-tuning leverages pre-trained weights while adapting the model to the specific dataset, saving time and computational resources compared to training from scratch. Fig. 3. Open in a new tab The Scaling of EfficientNet mode (Elmoufidi and Amoun 2021 ). a a baseline network; b – d conventional scaling that increases just one dimension of network width, depth, or resolution. e compound scaling method that uniformly scales all three dimensions with a fixed ratio The EfficientNet architectures family was created to enhance model performance, by determining the best way to scale convolutional neural networks (CNN). The model suggests a compound scaling method that uses a given set of coefficients to scale width, depth, and resolution equally. The EfficientNet-B0 CNN architecture is developed using this technique. Eight models, from B0 to B7, make up the EfficientNet model group; higher model numbers correspond to variants with more parameters and better accuracy (Hussain et al. 2023 ). By strengthening the network, CNNs are able to capture features that are more complex and rich (Bossard et al. 2014 ). But this can also result in the vanishing gradient issue, which increases the difficulty of network training. The network width adjustment allows it to collect more fine-grained features, which facilitates easier training. The EfficientNet-B0 model, depicted in Fig. 3 , accepts input images of size 224 × 224 × 3, where 224 × 224 represents the width and height of the image, and 3 signifies the image channels. This architecture captures features across layers using a mobile inverted bottleneck convolution and multiple convolutional layers with a 3 × 3 receptive field. The following equations demonstrate the scaling of width, depth, and resolution: w = β ϕ , 1 d = α ϕ , 2 r = γ ϕ , 3 s . t α . β 2 . γ 2 ≈ 2 4 α ≥ 1 , β ≥ 1 , γ ≥ 1 . 5 where the constant coefficients α, β, and γ are found via a small grid search on the initial tiny model, and w, d, and r stand for width, height, and resolution, respectively. EfficientNet uses a compound coefficient to equally scale the network width, depth, and resolution. However, high-level features are beyond the reach of vast and shallow networks. CNNs can detect finer patterns in high resolution images, and processing larger images requires more computing power and memory. Because of this, we assessed their performance on the American Sign Language dataset in this work using the Efficientnet-B0 model. Furthermore, as EfficientNet reduces compute costs, battery consumption, and training and inference speeds, it is very helpful for using deep learning on the edge (Chaitanya et al. 2023 ). Table 1 provides a summary of the fine-tuning applied to the EfficientNet-B0 model for gluten detection. The model comprises the EfficientNet-B0 backbone, which outputs a shape of (None, 1280) and consists of 4,019,812 parameters. Following this, a Dropout layer is included, which does not add any additional parameters. The final dense layer outputs a shape of (None, 2) and adds 2562 parameters. The model has 4,022,374 total parameters, all of which may be trained; there are no non-trainable parameters. This indicates that all parameters in the model are adjusted during the training process. Table 1. An overview of the fine-tuning applied to the EfficientNet-B0 model Layer (type) Output Shape Param EfficientNet-B0 (Functional) (None, 1280) 4,019,812 Dropout (None, 1280) 0 dense (Dense) (None, 2) 2562 Total params 4,022,374 Trainable params 4,022,374 Non-trainable params 0 Open in a new tab Results This section describes the results of the gluten and non-gluten food image classification model, including the accuracy curve, loss curve, confusion matrix, and classification table. After completing the training, the model achieved significant accuracy improvements. Specifically, the model attained a training accuracy of 99.02% and a validation accuracy of 98.38%, as indicated by the training progress output. The number of epochs for training and validation on gluten and non-gluten food images are 20. The graph in The accuracy and losses of the model with 20 classes are shown in Fig. 4 . The graph demonstrates the gap between training and validation accuracy decreased as the number of epochs increased, with the final difference was about 5% at end, suggesting only slight overfitting of the data. Fig. 4. Open in a new tab Model’s accuracy and loss curve The confusion matrix shows in Fig. 5 a. Shows the number of true and false predictions for every class. For the “Gluten” class, the model correctly predicted 170 instances (true positives) and incorrectly classified 50 instances as “NonGluten” (false negatives). For the “NonGluten” class, the model correctly predicted 117 instances (true negatives) but incorrectly classified 63 instances as “Gluten” (false positives). In our analysis of the confusion matrix (Fig. 5 ), we observed certain misclassifications between the “Gluten” and “NonGluten” categories. Specifically, 5 images of gluten-containing foods were predicted as “NonGluten,” and 10 images of non-gluten foods were predicted as “Gluten.” The misclassified items often shared visual characteristics, such as texture or color, which may have led to confusion in the model’s predictions. For instance, gluten-free and gluten-containing baked goods might appear similar. Similar features in ingredients that aren’t visibly distinguishable could have contributed to misclassifications, particularly in foods that have similar presentation styles. Fig. 5. Open in a new tab a Confusion Matrix and b Comparison of test accuracies across different models on Food101 dataset The classification report for the gluten image classification model using the EfficientNet-B0 pretrained model is generated. The classification report reveals that the model has an overall accuracy of 99%. For the “Gluten” class, the model achieved a precision of 0.99, recall of 0.99, and an F1-score of 0.99, while for the “Non-Gluten” class, it obtained a precision of 0.98, recall of 0.97, and an F1-score of 0.98. The macro average precision, recall, and F1-score are 0.99, 0.98, and 0.98, respectively, with a weighted average of 0.99 for all three metrics. The model demonstrates excellent performance across both classes, showing high precision, recall, and F1-scores, indicating strong and reliable classification capabilities. The comprehensive evaluation through these metrics underscores the model’s robustness and effectiveness in gluten detection. The high accuracy with balanced precision and recall indicate that the model can reliably classify food images, making it a significant tool for dietary management and with allergen detection. The comparative study of accuracy across different models using the Food101 dataset reveals the superiority of the proposed model in this study. Figure 5 b shows the comparison of test accuracies across different models on Food101 dataset. Traditional approaches such as Random Forest (Islam et al. 2018 ) achieved a relatively low accuracy of 50.76%, while transfer learning techniques have shown improvements. For instance, using Inception-ResNet and Inception V3 models, accuracies of 72.55% (Top-1) and 91.31% (Top-5) were achieved (Uddin et al. 2021 ). The EfficientNetB0 model previously attained an accuracy of 80% (Abiyev and Adepoju 2024 ), and another approach that combined training from scratch with transfer learning yielded 61.4% (Top-1) and 85.2% (Top-5) (Yanai and Kawano 2015 ). The extended Food-101 dataset, focused on sub-continental foods, reached a Top-1 accuracy of 85% using deep learning and semantic analysis (Memiş et al. 2020 ). However, the proposed model in this study, which fine-tunes the EfficientNetB0 with early stopping and leverages a curated Food101 dataset with 20 classes, outperformed all previous models, attaining an impressive accuracy of 99%. The improvement is significant and underscores the effectiveness of the methodological enhancements introduced in this work. The comparative analysis of the test accuracies demonstrates that the proposed model using transfer learning with early stopping on a curated Food101 dataset significantly outperforms other models, accomplishing a remarkable accuracy of 99%, which is highest among all the approaches considered. This highlights the effectiveness of the tailored dataset and the methodological refinements in enhancing model performance for gluten detection in food images. Conclusion and future work Accurate identification of food categories is crucial for intelligent health management and has garnered considerable research interest in recent years. Leading benchmarks like Food101 offer extensive food image datasets that drive advancements in this area of study. The convolutional neural network approach used in the proposed framework performs exceptionally well, provided there is a sufficiently large dataset for training. Training the model on such an extensive dataset can take hours or even days, which is why transfer learning is employed in this approach. The goal of the work is to improve the model’s accuracy in classification of gluten and non-gluten food images, aiming to encompass nearly all major cuisines consumed worldwide. However, a significant challenge encountered during this research was the scarcity of GPU resources, even while using Google Colab, which limited the speed and efficiency of the model training process. It is quite feasible to advance further using a CNN with highly dense connections, which can learn features faster and more effectively than the existing method. Artificial Intelligence can help fine-tuning the pre-trained EfficientNet model to almost perfect, as seen in, achieving an almost ideal framework for image classification. In the future, this work can be expanded by incorporating more diverse datasets from different cuisines to improve the model’s robustness. Additionally, developing a mobile application that utilizes the phone camera along with these enhanced features would provide a practical, one-stop solution where users can maintain their nutrition and overall health. Authors’ contributions Mayura Tapkire: Review of Literature, Planning, execution, Manuscript draft writing. Vanishri Arun: Planning, Supervision and manuscript validation. Lavanya M S: Execution, manuscript draft. Shashidhar R: Review, Manuscript draft writing. Funding No funding. Data availability Data transparency will be made available on request. Code availability Software application or custom code will be avail. Declarations Conflict of interests The author declares that she has no conflict of interest. Ethics approval Not applicable. Consent to participate All the authors give consent to be part of the manuscript submission as authors. Consent for publication All authors give consent for the publication of the manuscript along with the data submitted. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Abiyev R, Adepoju J (2024) Automatic food recognition using deep convolutional neural networks with self-attention mechanism. Human-Centric Intell Syst 4(1):171–186 [ Google Scholar ] Ao S, Ling CX (2015) Adapting new categories for food recognition with deep representation. In: 2015 IEEE International conference on data mining workshop (ICDMW). IEEE, pp 1196–1203 Arun V, Lavanya MS, Tapkire M (2024) HML-PCD: A Hybrid machine learning technique for Early prediction and classification of celiac disease. Int J Intell Syst Appl Eng 12(3):1338–1346 [ Google Scholar ] Asenjo Ponce de León M (2020) Applying deep learning for food image analysis (Master’s thesis, Universitat Politècnica de Catalunya) Bahadur S, Kushwaha R, Sharma S, Anuradha SG (2022) Using deep learning techniques, a framework for estimating the nutritional value of food in real time. Int J Health Sci II:7635–7648 [ Google Scholar ] Bossard L, Guillaumin M, Van Gool L (2014) Food-101–mining discriminative components with random forests. In: Computer vision–ECCV 2014: 13th European conference, Zurich, Switzerland, September 6–12, 2014, proceedings, part VI 13. Springer International Publishing, pp 446–461 Chaitanya A, Shetty J, Chiplunkar P (2023) Food image classification and data extraction using convolutional neural network and web crawlers. Procedia Comput Sci 218:143–152 [ Google Scholar ] Elmoufidi A, Amoun H (2021) Efficientnetb3 architecture for diabetic retinopathy assessment using fundus images Giovany S, Putra A, Hariawan AS, Wulandhari LA (2017) Machine learning and sift approach for Indonesian food image recognition. Procedia Comput Sci 116:612–620 [ Google Scholar ] Hinton G, Deng L, Yu D et al (2012) Deep neural networks for acoustic modeling in speech recognition: the shared views of four research groups. IEEE Signal Process Mag 29:82–97. 10.1109/MSP.2012.2205597 [ Google Scholar ] Hussain A, Ul Amin S, Fayaz M, Seo S (2023) An Efficient and robust hand gesture recognition system of sign language employing finetuned inception-V3 and efficientnet-B0 network. Comput Syst Sci Eng 46(3):3509–3525 [ Google Scholar ] Islam MT, Siddique BNK, Rahman S, Jabid T (2018) Food image classification with convolutional neural network. In: 2018 international conference on intelligent informatics and biomedical sciences (ICIIBMS), vol 3. IEEE, pp 257–262 Jiang M (2019) Food image classification with convolutional neural networks. CS230: Deep Learning, Fall Krizhevsky A, Sutskever I, Hinton GE (2017) ImageNet classifcation with deep convolutional neural networks. Commun ACM 60:84–90. 10.1145/3065386 [ Google Scholar ] LeCun Y, Bengio Y, Hinton G (2015) Deep learning. Nature 521:436–444. 10.1038/nature14539 [ DOI ] [ PubMed ] [ Google Scholar ] Memiş S, Arslan B, Batur OZ, Sönmez EB (2020. A comparative study of deep learning methods on food classification problem. In: 2020 Innovations in intelligent systems and applications conference (ASYU). IEEE, pp 1–4 Pan L, Pouyanfar S, Chen H, Qin J, Chen SC (2017) Deepfood: Automatic multi-class classification of food ingredients using deep learning. In: 2017 IEEE 3rd international conference on collaboration and internet computing (CIC). IEEE, pp.181–189 Park SJ, Palvanov A, Lee CH, Jeong N, Cho YI, Lee HJ (2019) The development of food image detection and recognition model of Korean food for mobile dietary management. Nurs Res Pract 13(6):521–528 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Stoleru CA, Dulf EH, Ciobanu L (2022) Automated detection of celiac disease using Machine Learning Algorithms. Sci Rep 12(1):4071 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Tapkire MD, Arun V (2022) A Survey on celiac disease prediction using AI Techniques. Int J Adv Res Comput Commun Eng. 10.17148/IJARCCE.2022.11346 [ Google Scholar ] Tapkire MD, Arun V (2023) Application of artificial intelligence to corelate food formulations to disease risk prediction: a comprehensive review. J Food Sci Technol 60:2350–2357. 10.1007/s13197-022-05550-w [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Tompson J, Jain A, LeCun Y, Bregler C (2014) Joint training of a convolutional network and a graphical model for human pose estimation. In: Proceedings of the 27th international conference on neural information processing systems, vol 1. MIT Press, Cambridge, pp 1799–1807 Uddin AM, Al Miraj A, Sarma MS, Das A, & Gani MM (2021) Traditional Bengali food classification using convolutional neural network. In: 2021 IEEE region 10 symposium (TENSYMP). IEEE, pp 1–8 VijayaKumari G, Vutkur P, Vishwanath P (2022) Food classification using transfer learning technique. Glob Trans Proc 3(1):225–229 [ Google Scholar ] Yanai K, Kawano Y (2015) Food image recognition using deep convolutional network with pre-training and fine-tuning. In: 2015 IEEE international conference on multimedia & expo workshops (ICMEW). IEEE, pp 1–6 Yu Q, Mao D, Wang J (2016) Deep learning based food recognition. Technical report, Stanford University Zammit SC, McAlindon ME, Greenblatt E, Maker M, Siegelman J, Leffler DA, Sidhu R (2023) Quantification of celiac disease severity using video capsule endoscopy: a comparison of human experts and machine learning algorithms. Current Med Imag 19(12):1455 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Zong Z, Nguyen DT, Ogunbona P, Li W (2010) On the combination of local texture and global structure for food classification. In: 2010 IEEE International symposium on multimedia. IEEE, pp 204–211 Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Availability Statement Data transparency will be made available on request. Software application or custom code will be avail. Articles from Journal of Food Science and Technology are provided here courtesy of Springer ACTIONS View on publisher site PDF (1.2 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Advances in Sensing Techniques and Deep Learning for Food Quality Detection: Opportunities, Challenges, and Perspectives from Image Recognition to Multimodal Data Fusion.#1091525
• Deep learning model for real time moisture content detection and prediction in white tea withering using near infrared spectroscopy.#610347
• DeepFixel: Crossing white matter fiber identification through spherical convolutional neural networks.#224883
• Lemon Classification Using Deep Learning#116500
• Plastic Waste Classification Using Deep Learning: Insights from the WaDaBa Dataset#972897