Food Recipe Recommendation Based on Ingredients Detection

Using Deep Learning

Md. Shafaat Jamil Rokon, Md Kishor Morol, Ishra Binte Hasan, A. M. Saif, Rafid Hussain Khan
Department of Computer Science
American International University-Bangladesh (AIUB)
{shafaatjamilrokon, kishoremorol, ishrahasan12, alimdsaif1996, rafidkhan321.rk}@gmail.com

ABSTRACT
Food is essential for human survival, and people always try to 1 INTRODUCTION
taste different types of delicious recipes. Frequently, people People nowadays become very much health conscious and,
choose food ingredients without even knowing their names or they try to take at most healthy food in their meal. An appropriate
pick up some food ingredients that aren’t obvious to them from a food recipe must require serving the real taste of the food and a
grocery store. Knowing which ingredients can be mixed to make a balanced diet, and healthy life. As we all know, health is wealth,
delicious food recipe is essential. Selecting the right recipe by so food selection is essential for better health. Sometimes people
choosing a list of ingredients is very difficult for a beginner cook. pick some ingredient that they do not even know their names. Or
However, it can be a problem even for experts. There is the pick up some ingredients from a grocery store, but they do not
constant use of machine learning in our everyday lives. One such know how to make recipes using those ingredients. People should
example is recognizing objects through image processing. know which ingredients can be mixed and make delicious recipes.
Although this process is complex due to different food For a beginner, it is very challenging to select a recipe just by
ingredients, traditional approaches will lead to an inaccuracy rate. seeing the ingredients. Even for an expert cook, it is also
These problems can be solved by machine learning and deep problematic. In this paper, we try to solve this problem. We
learning approaches. In this paper, we implemented a model for implemented an approach of recognizing food ingredients, and
food ingredients recognition and designed an algorithm for after analyzing and classifying these recognized ingredients, food
recommending recipes based on recognized ingredients. We made that can be cooked with compatible recipes will be recommended.
a custom dataset consisting of 9856 images belonging to 32 This recommendation will help people to get a suitable food
different food ingredients classes. Convolution Neural Network recipe. Online cooking recipe sites have become very popular, but
(CNN) model was used to identify food ingredients, and for recipe they sometimes get confused when people want to cook by
recommendations, we have used machine learning. We achieved following these sites. This may cause to stop users from referring
an accuracy of 94%, which is quite impressive. to these types of cooking sites throughout searching at grocery
stores. Technology has to be improved to solve these problems,
and then object recognition technology was introduced to the
world. There has been significant progress in object recognition
technology. Object recognition is a technology that identifies the
objects shown in an image based on their specific properties [1].
So, Object recognition is a process that finds things in the real
world from analyzing an image. To use object recognition,
KEYWORDS
computer vision, and deep learning models, open-source software
libraries such as the Open Computer Vision Library (OpenCV)
Deep Learning, CNN, Ingredients Detection, [2], TensorFlow [3], NumPy [4], and Keras [5] have been used
Recipe Recommendation, Resnet50 widely. With these libraries, we can implement object recognition
schemes on PCs and mobile devices such as iPhones and Android
ACM Reference format:
smartphones. With the recent progress of mobile devices and their
explosive spread, we have now been able to recognize objects at
Md. Shafaat Jamil Rokon, Md Kishor Morol, Ishra Binte Hasan, A. M.
Saif, Rafid Hussain Khan. 2022. Food Recipe Recommendation Based on any time on a mobile device. OpenCV [2] can be used in image
Ingredients Detection Using Deep Learning. In International Conference processing and representation. Deep learning APIs like NumPy
on Computing Advancements (ICCA 2022), March 10–12, 2022, Dhaka, can do the numerical computation, making machine learning
Bangladesh.
faster and easier. TensorFlow[3] API can be used to recognize the
image and train deep learning architectures for image
classification [4]. There are pre-trained deep learning architectures
such as MobileNet [6], ResNet [7], and Inception [8], VGG16 [9] nearest neighbors to recommend recipes from the Recipe 1M
can be used too. The advantage of using pre-trained architecture is dataset.
that we can modify it using transfer learning [10]. With the help Chen, J. et al. [17] used a dataset composed of 61,139 image-
of these libraries, APIs, and architectures, we can easily recognize recipe pairs gathered from the Go Cooking website. They have
objects from an image. explored the recent advances in cross-modality learning to address
the problems as mentioned earlier. A deep model stacked attention
For recipes identification, very little work has been done network (SAN) is absorbed and revised in their system.
because of not having a proper dataset, and also, working on
multiple object recognizing is still challenging. Most of the papers KeijiYanai, et al. [18] made a food ingredient recognition system
we found are just on recipes identification or food for android OS. The system used a color-histogram-based bag-of-
recommendation. But we aim to detect food ingredients by features pull out from several frames as an image file and a linear
reaching out maximum classification rate and recommending kernel SVM as a classifier. They stored 30 videos in the database,
worthy food recipes. In the following section we discuss about and their accuracy was 83.93%.
related previous works.
Suyash Maheshwari et al. [19] proposed two algorithms to
recommend ingredients. Cooking and tasting experiments showed
the proposed methods were effective for each purpose. Through
2 RELATED WORKS cooking demonstration with the recommended alternative
Food detection, ingredient detection, or recipe ingredients and subjective evaluation experiments, it was
recommendation have received increasing attention in recent confirmed that both algorithms recommended acceptable
years. All these profound deep learning-based works related to ingredients for over 90 percent. However, algorithm one might
object detection and classification. There are adequate deep not recommend an alternative ingredient similar to the exchange
learning models and methods available for object detection and ingredient.
classification. Use of color histogram, BoF, linear kernel SVM
classifier, and K-nearest neighbors are well known. Mona Mishra, Yifan Gong et al. [20] for their work, thousands of
images are trained for various categories to build a CNN for
Kawano et al. [12] presented an approach for a real-time food image recognition. And predict what the object in the image from
recognition system for smartphones. Two types of real-time image users upload.
recognition methods have been used. One is the combination of
bag-of-features (BoF) and color histograms with X2 kernel feature So, we understood that we have a good scope of making a model
maps. And the second one is the HOG patch descriptor and a that can identify ingredients and then recommend recipes based
color patch descriptor with the state- of-the-art Fisher Vector on those ingredients. In the following section, we will discuss our
representation. Linear SVM has been used as a classifier. They proposed model.
have achieved an accuracy of 79.2% classification rate.

Bolaños et al. [13], used CNN for food image recognition. They
use two different inputs for their method. The first one provides a 3 METHODOLOGY
low-level description of the food image. The penultimate layer of
the InceptionResNetV2 CNN is used in the first method, and the 3.1 Proposed Method
second one provides a high-level description of the food image First, we divided our work into two parts. One is identifying 32
using LogMeals API. Three different CNNs were supplied by this ingredients where we used Convolution Neural Network (CNN),
LogMeals API that predicts food groups, ingredients, and dishes. and after getting the result, we proceed to the second part. Secord
part is to the recipe recommendations containing 19 different
Chang Liu et al. [14], proposed CNN based novel approach for classes.
visual-based food image recognition with 7-layer architecture and
achieved 93.7% top-5 accuracy using the UEC-100 dataset were We have used Transfer Learning for training our CNN model for
existing approaches like SURF- BoF + Color Histogram and the first part. Reuse of a pre-trained model is known as transfer
MKL could achieve 68.3% to 76.8% top-5 accuracy with the same learning. As a pre-trained model, we have used ResNet50.
dataset. Transfer learning is immensely popular in deep learning because
it can train deep neural networks with relatively little data [11].
Raboy-McGowan and L. Gonzalez, et al. [15][16] both used the Transfer learning is efficient since most real-world problems don't
Recipe 1M dataset and introduced Recipe Net, a food to recipe have millions of labeled data points to train such complex deep
generator trained on the Recipe 1M dataset. They have used neural networks.
ResNet-50, DenseNet-121 convolutional neural networks tclassify
food images and encode their features. After that, they used K- As a base model, we have used 50 layers deep convolution neural
network, which is ResNet50. ResNet50 was trained on more than
a million images in the ImageNet dataset. The pre-trained network flatten layer and a new prediction layer for our required 32
can classify images into 1000 categories, but we need 32 object classes.
categories such as Fried Rice, Chotpoti, Subway, etc. Pre-trained
ResNet50 does not classify images by those 32 specific
categories. We created a new model from scratch for this specific
purpose, but for good results, we would need many images with
labels for determining Fried Rice and Chotpoti, etc.

Fig. 2. Block diagram for proposed Transfer Learning model

Some layers before that in the pre-trained model may identify

features like roads, buildings, windows, etc. will drop in a
replacement for the last layer of ResNet50 model, this new last
layer will predict whether an image is Chotpoti or Subway or any
other from 30 classes.

Fig. 1. Block diagram of the proposed method

It was possible to gain good results with fewer data using transfer Fig. 3. Connection between last remaining layer and new prediction layer
learning. That's why we chose transfer learning. An early layer of
a deep learning model identifies shapes, a last layer identifies
more complex visual patterns, and the final layer makes We have a lot of connections here. In figure 3, we can see the
predictions. Due to the similar low-level patterns involved in most ResNet50 model has many layers, we have just cut off the last
computer vision problems, most layer layers of a pre-trained layer. What's left in the last layer is information about our image
model are helpful for new applications. This means that most of content stored as a series of numbers in a tensor. It should be a
the layers of the pre-trained ResNet50 models can be reused and one-dimensional tensor, also known as a vector. It can be shown
we need to replace only that final layer that is used to generate as a series of dots. Dots are called nodes. The first node represents
predictions. That is why we've cut down the last prediction layer the first number in the vector and the second node represents the
of the ResNet50 model trained for 1000 classes and added a
second number and so on. We want to classify the images into 32 Our next step is to use an average pool, and then finish it up
categories: Beef Meat, Bell pepper, Burger Bun, Burger Patty, with a fully connected layer of 1000 nodes. Lastly, there is a
Chickpea, Cheese, Chicken, Chocolate, Chocolate Syrup, Donut SoftMax function that provides us 1 layer. Max-pooling layers,
Bun, Egg, Ground Beef, Hotdog Bun, Ice Cream, Lasagna along with activation functions, are not counted. In total, we have
Noodles, Letus, Mayonnaise, Milk, Mushrooms, Nachos Chips, a 1+9+12+18+9+1=50 layers Deep Convolutional Network.
Noodles, Onion, Pasta, Pizza Dough, Potato, Rice, Roll Bread,
Sandwich Bread, Sausage, Shawarma Bread, Subway Bun, and
Tomato. In the last layer, we keep the pre-trained model and add 3.3 Recipe Recommendation
another layer with 32 nodes. The first node capture how the Beef
Meat image is, the second node capture how the Bell pepper As mentioned before our aim is to recommend cooking recipes
image is, the third node capture how the Burger Bun image is, and from recognized food ingredients. We have generated our recipe
so on. We've used training data to determine which nodes suggest database and algorithm for recipe recommendations. We have
an image is Burger Bun, which is Pasta, and so on. selected 19 kinds of cooking recipes associated with 32
ingredients and designed a 2D matrix (19 rows × 32 columns) for
3.2 Architecture the recipe recommendation algorithm in which 19 rows contain 19
It is a variant of ResNet [7]. In ResNet 50 there are 48 recipes and 32 columns contain 32 food ingredients.
convolution layers as well as 1 MaxPool and 1 Average Pool
layer. In this network, there is a total of 50 layers

• The first layer will have a feature size of 7*7 and 64

such filters all with a stride of size 2. And it gives 1
layer.
• Next, we will see max pooling with the window size of
3*3 and a stride size of 2.
• 1 * 1, 64 means the first convolution layer has a filter
size of 1 * 1, and 64 such filters. Next 3 * 3, 64 means
filter size of 3 * 3, and 64 such filters and at last 1 * 1,
256 means, filter size of 1 * 1, and 256 such filters.
These three layers are repeated 3 times which means 9
layers in this step.

Fig. 5. 2D matrix (19 rows × 32 columns) for recipe recommendation

algorithm.

In figure 5, we created a one-to-many relationship with food

ingredients and recipes. We have objectified here which recipes
can be cooked with which ingredients. For example, Egg is
needed for cooking Chotpoti, Fried Rice, and Noodles. So we
have assigned "1" in those rows where recipes can be cooked with
Egg and the rest of the rows remain "0" for Egg. It is like binary
Fig. 4. Architecture of ResNet50 classification. Chicken is needed for the highest number of
recipes. Nine recipes can be cooked with Chicken, those are
• After that1*1,128 layer, then 3*3,128, and at the last of Chicken Fry, Chicken Wings, Fried Rice, Noodles, Pasta, Roll,
1 * 1, 512 this step will be repeated 4 times so that gives Sandwich, Shawarma, and Subway. So we have assigned "1" in
us 12 layers in this step. these recipe rows against Chicken and the other ten rows for
• Next 1*1, 256 layers, then 3*3, 256, and then 1*1, recipes remain "0" against Chicken.
1024, and this will be repeated 6 times so that gives us
18 layers. After food ingredients recognition using our developed CNN
• And then again a1*1, 512 layers, 3*3, 512 and 1* 1, model, we've run a linear search in this database (19 × 32
matrices). For example, our developed CNN model recognized
2048, and this was repeated 3 times so that it gives 9
three food ingredients which are Chicken, Egg, and Rice. So our
layers.
developed algorithm will conduct a linear search in this database
and cross-check which recipes are assigned "1" for Chicken, Egg,
and Rice food ingredients columns and other food ingredients
columns are assigned “0”. Then it will find Fried Rice. Multiple 3.5 Data Augmentation
recipes can be recommended by our developed algorithm. For
example, the recognized food ingredients are Chickpea, Egg, and A machine learning model performs better and is more accurate
Potato. First of all the recipe recommendation algorithm will when the dataset is affluent and adequate. We’ve got pre-made
conduct a linear search in the database for Chickpea, Egg, and datasets of onion, tomato, potato, and rice among all the
Potato individually and cross- check which recipes are assigned ingredients. We’ve used data augmentation to increase the number
"1" only for Chickpea or Egg or Potato and then cross-check of photos for Beef Meat, Bell Pepper, Burger Bun, Burger Patty,
which recipes are assigned "1" for Chickpea, Egg and Potato Chickpea, Cheese, Chicken, Chocolate Syrup, Donut Bun, Ground
altogether. Then it will find Hotpot, French Fries, and Potato Beef, Hotdog Bun, Lasagna Noodles, Lettuce, Mayonnaise, Milk,
Wedges. This is how our developed recipe recommendation Mushrooms, Nachos Chips, Noodles, Pasta, Egg, Pizza Dough,
algorithm works.
Chocolate, Roll Bread, Sandwich Bread, Sausage, Shawarma
Bread, and Subway Bun. We used the TensorFlow library to
3.3 Dataset perform data augmentation. The Image Data Generator function
We had to create our own since there were no real-life datasets was used to supplement data that is a Keras function. During data
available online for this topic, such as pizza dough, lettuce, augmentation, images have been rotated 45 degrees, shifted width
mayonnaise, etc. For this study, we have used three datasets from and height by 20%, zoomed by 20%, flipped horizontally, and
Kaggle like food101, fruit 360, and UECFOOD256 and the data sheared with a range of 20%. Image data augmentation is usually
we created. From fruit360, we took the potato and tomato applied only to the training dataset, not the validation or
datasets, and from food101, we took the ice cream and onion assessment datasets.
datasets, and finally, from UECFOOD256, we took the rice,
chicken, roll bread, and sausage datasets. In total, we made the
dataset which contains a total of 9856 images of 32 different 3.6 Optimizer and Learning Rate
ingredients in figure 6.
In optimization, optimizers are algorithms or methods that modify
deep neural network attributes, such as weights and learning rates,
to reduce losses. Optimizers are used to minimize a function as
part of solving optimization problems. There are various kinds of
optimizers available as like RMSprop, Adam, Nadam, SGD, Ftrl
etc. Adam optimizer [21] is used in our research. Adam
optimization is a technique that extends stochastic gradient
descent for use in computer vision and natural language
processing. It is commonly used by researchers in PC vision
because of its better performance. In our model Adam optimizer is
used with the learning rate of 0.001. It’s the default learning rate
of Adam optimizer. In training a convolution neural network, the
learning rate plays a significant role. A lower learning rate will
yield better classification; however, the optimizer will set aside
more effort to achieve global optima, which will reduce the loss.
The higher the learning rate, the worse the accuracy. 0.001 was set
Fig. 6. 32 kinds of food ingredients in our dataset as the learning rate as that would naturally drop upon checking the
validation accuracy. For calculating the loss we used the
3.4 Preprocessing Categorical cross- entropy function. Categorical cross-entropy is
also known as Softmax loss. This is a Softmax activation with a
Since convolution neural networks accept inputs of the same size, cross- entropy loss as well. In comparison to other functions, like
all the images of food ingredients should be resized to a discerned classification errors, mean squared errors, etc., we found that
size before inputting them to the CNN. The targeted input size for cross-entropy shows a good quality [22] since our model is a
ResNet50 was 224 × 224. But our dataset images were in different multiclass classification problem, Categorical cross-entropy is the
shapes and sizes as we gathered them from different sources. fittest choice for us.
Some food ingredients datasets were available free on Kaggle and
other datasets we created our own by a camera. Training a CNN
on unreformed images will surely lead to terrific classification
exhibitions. So we have resized the images into 224 × 224 pixels
and excluded unnecessary objects from the images.
4 RESULTS AND DISCUSSION 4.2 Result Comparison
4.1 Model Performance Research for recipe recommendations can hardly be seen. There
As we mentioned, we have a dataset of a total of 9856 images. So, are some papers on food item recognition that are roughly like our
we took 70% of the images for the training, 20% for testing, and study.
the rest of the 10% for validation. After running 20 epochs using
transfer learning, our proposed CNN-based ResNet50 model TABLE 1. COMPARISON BETWEEN SOME PREVIOUS
achieved a 99.71% accuracy for the training dataset. And 92.6% WORKS
accuracy on the validation dataset. After analyzing the results and
confusion matrix, it can be stated that the CNN model has Work Technology Accuracy
acceptable performance for classifying food ingredients. The used
CNN model's performance is shown below in figure 7 and 8. Khan, et al. “A machine learning CNN 90.47%
approach to recognize junk food
(proposed Model)” [23]
Sun, et al. “Yelp Food Identification CNN 68.49%
via Image Feature Extraction and &SVM
Classification” [24]
Yanai, et al. “Food image recognition DCNN 70.4%
using deep convolutional network
with pre- training and fine tuning”
[25]
Baxter, et al. “Food recognition using CNN 81.2%
ingredient-level features” [26]
R. Xu, et al “Modeling restaurant CNN 84.9%
context for food recognition” [27]
Fig. 7. Training and Validation accuracy. Ciocca, et al. “Food recognition: a CNN 79%
new dataset, experiments, and
results” [28]
Mayers, et al. “Im2calories: towards Googlre Net 79%
an automated mobile vision food
diary” [29]
Recipe recommendation based on CNN-based 94%
ingredients detection using deep Resnet50
learning

4 CONCLUSION

This paper presented a CNN model to recognize food

Fig. 8. Training and Validation loss ingredients and a recipe recommendation algorithm based on
detected ingredients to suggest cooking recipes. In addition, we
introduced a custom dataset of 32 categories of food ingredients.
We achieved the testing accuracy of 94%, which is impressive,
and proved that the performance of this model to recognize food
ingredients from images is more advanced. Due to these high
accuracy levels, we found CNN to be particularly suitable for
food ingredients recognition. Moreover, if we could get a high
configuration of computer GPU, then the accuracy level could be
above the current accuracy level as mentioned before; we used
transfer learning for training the CNN model. ResNet50 was used
as a base model. As part of an experiment, we also used VGG16
and MobileNetV2 as a base model, but ResNet50 generated better
performance than the other two pre-trained models. We selected [20] Mishra, M. and Y. Gong, Recipe Recommender System Using Image
Recognition of Food Ingredients. 2018.
19 cooking recipes against 32 food ingredients and developed a [21] Kingma, D.P. and J. Ba, Adam: A method for stochastic optimization. arXiv
unique algorithm for recipe recommendations based on preprint arXiv:1412.6980, 2014.
[22] Janocha, K. and W.M. Czarnecki, On loss functions for deep neural networks in
ingredients detection. classification. arXiv preprint arXiv:1702.05659, 2017.
In the future, we aim to increase the number of recipes and food [23] Khan, T.A., et al. A machine learning approach to recognize junk food. in 2019
ingredients and enrich our dataset. We’ve found two problems in 10th International Conference on Computing, Communication and Networking
Technologies (Icccnt). 2019. IEEE.
our dataset; one is contrast variation of images, and another is [24] Sun, F., Z. Gu, and B. Feng, Yelp Food Identification via Image Feature
imbalance problem of classes. We used data augmentation to Extraction and Classification. arXiv preprint arXiv:1902.05413, 2019.
[25] Yanai, K. and Y. Kawano. Food image recognition using deep convolutional
reduce the imbalance problem. In the future, we will work to network with pre-training and fine- tuning. in 2015 IEEE International
overcome these two problems completely. We have got a Conference on Multimedia & Expo Workshops (ICMEW). 2015. IEEE
limitation in our research: our CNN model cannot classify [26] Baxter, J., Food recognition using ingredient-level features. Ηλεκτρονικό].
Available: http://jaybaxter . net/6869_food_project. pdf, 2012.
multiple objects. That means our CNN model can recognize a [27] Herranz, L., S. Jiang, and R. Xu, Modeling restaurant context for food
single food ingredient at a time. It cannot identify various food recognition. IEEE Transactions on Multimedia, 2016. 19(2): p. 430-440.
[28] Ciocca, G., P. Napoletano, and R. Schettini, Food recognition: a new dataset,
ingredients from a given image. We want to deploy multiple experiments, and results. IEEE journal of biomedical and health informatics,
object detection in our research in the future. 2016. 21(3): p. 588-598.
[29] Meyers, A., et al. Im2Calories: towards an automated mobile vision food diary.
in Proceedings of the IEEE International Conference on Computer Vision.
2015.
REFERENCES [30] S. I. Sohel, C. Mondol, H. S. Ayon, U. T. Islam and M. K. Morol, "Music
Suggestions from Determining the Atmosphere of Images," 2021 24th
[1] Bergevin, R. and M.D. Levine, Generic object recognition: Building and International Conference on Computer and Information Technology (ICCIT),
matching coarse descriptions from line drawings. IEEE Transactions on Pattern 2021, pp. 1-7, doi: 10.1109/ICCIT54785.2021.9689781.
Analysis and Machine Intelligence, 1993. 15(1): p. 19-36.
[2] Bradski, G. and A. Kaehler, OpenCV. Dr. Dobb’s journal of software tools,
2000. 3.
[3] Abadi, M., et al. Tensorflow: A system for large-scale machine learning. in 12th
{USENIX} symposium on operating systems design and implementation
({OSDI} 16). 2016.
[4] McKinney, W., Python for data analysis: Data wrangling with Pandas, NumPy,
and IPython. 2012: " O'Reilly Media, Inc.".
[5] Ketkar, N., Introduction to keras, in Deep learning with Python. 2017, Springer.
p. 97-111.
[6] Sinha, D. and M. El-Sharkawy. Thin mobilenet: An enhanced mobilenet
architecture. in 2019 IEEE 10th Annual Ubiquitous Computing, Electronics &
Mobile Communication Conference (UEMCON). 2019. IEEE.
[7] Targ, S., D. Almeida, and K. Lyman, Resnet in resnet: Generalizing residual
architectures. arXiv preprint arXiv:1603.08029, 2016.
[8] Szegedy, C., et al. Rethinking the inception architecture for computer vision. in
Proceedings of the IEEE conference on computer vision and pattern
recognition. 2016.
[9] Qassim, H., A. Verma, and D. Feinzimer. Compressed residual-VGG16 CNN
model for big data places image recognition. in 2018 IEEE 8th Annual
Computing and Communication Workshop and Conference (CCWC). 2018.
IEEE.
[10] Pan, S.J. and Q. Yang, A survey on transfer learning. IEEE Transactions on
knowledge and data engineering, 2009. 22(10): p. 1345-1359.
[11] A. M. Faruk, H. A. Faraby, M. M. Azad, M. R. Fedous and M. K. Morol,
"Image to Bengali Caption Generation Using Deep CNN and Bidirectional
Gated Recurrent Unit," 2020 23rd International Conference on Computer and
Information Technology (ICCIT), 2020, pp. 1-6, doi:
10.1109/ICCIT51783.2020.9392697.
[12] Kawano, Y. and K. Yanai, Foodcam: A real-time food recognition system on a
smartphone. Multimedia Tools and Applications, 2015. 74(14): p. 5263-5287.
[13] Bolaños, M., M. Valdivia, and P. Radeva. Where and what am i eating? image-
based food menu recognition. in European Conference on Computer Vision.
2018. Springer.
[14] Liu, C., et al. Deepfood: Deep learning-based food image recognition for
computer-aided dietary assessment. in International Conference on Smart
Homes and Health Telematics. 2016. Springer.
[15] Raboy-McGowan, D., S. Lu, and L. Gonzalez, RecipeNet: Image to
Recipe/Nutritional Information Generator.
[16] Salvador, A., et al. Inverse cooking: Recipe generation from food images. in
Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern
Recognition. 2019.
[17] hen, J., L. Pang, and C.-W. Ngo. Cross-modal recipe retrieval: How to cook this
dish? in International Conference on Multimedia Modeling. 2017. Springer.
[18] Yanai, K., T. Maruyama, and Y. Kawano, A Cooking Recipe Recommendation
System with Visual Recognition of Food Ingredients. International Journal of
Interactive Mobile Technologies, 2014. 8(2).
[19] Maheshwari, S. and M. Chourey, Recipe Recommendation System using
Machine Learning Models. International Research Journal of Engineering and
Technology (IRJET), 2019. 6(9): p. 366-369.