As an avid "eater" (or foodie), I've always loved to travel to different cities and countries and explore different types of cuisines. Eating is by far one of the most exciting things I have done, other than studying data science, so I wanted to combine the two together. This is an ongoing project that I am still working on. At the current stage, I have only trained the model on 105 different Japanese food dishes, which can be found in this file. I am not Japanese, nor am I an expert at Japanese food. However, it is one of my favorite type of cuisines, which is why I chose this as my starting point. My hope is that users can use the web app easily by uploading an image of a dish and getting an accurate feedback of what the food is (in English). At the moment, it only provides the classification label, but I have plans of providing other details like nutritional facts in the future. I also plan on creating other models with other cuisines so that the web app can serve different types of cuisines.
If you just want to test it out, there is a version of the web app currently running here.
If you want to run it locally (perhaps with your own model), you can run the following code from the root folder:
cd web_app
pip install -r requirements.txt # to install necessary packages
python app.py
The first step for this project is to collect and clean the data. To start, I needed to know what types of Japanese food there were, so I went on a site that had 105 popular Japanese dishes and scraped all the names of the dishes there. The script can be found here. Running this script would create a .csv file of a diverse list of 105 Japanese dishes. This list was scraped from a website that listed popular Japanese dishes. Yes, I do understand that there are definitely more than 105 unique Japanese dishes, but I had to start somewhere. Because I will be using an ImageGenerator from Keras so that I could use the .flow_from_directory() method to augment my images, I needed my images folder structure to look like this:
images
├── train
│ ├── Nigiri
│ │ ├── Nigiri_001.jpg
│ │ ├── Nigiri_002.jpg
│ │ └── ...
│ ├── Edamame
│ │ ├── Edamame_001.jpg
│ │ ├── Edamame_002.jpg
│ │ └── ...
│ └── ...
├── valid
│ ├── Nigiri
│ │ ├── Nigiri_272.jpg
│ │ ├── Nigiri_.428jpg
│ │ └── ...
│ ├── Edamame
│ │ ├── Edamame_154.jpg
│ │ ├── Edamame_221.jpg
│ │ └── ...
│ └── ...
└── test
├── ...
After setting up my folder structure, I ran the script to download more than 500 images for each food item in my list. To get better quality images, I manually created a translated document in the food_list folder so that my search results could be in Japanese. To then split my data into train, valid, and test portions, I created a script to setup the images folder in preparation for my model.
Once I have collected all the data, I built my model. This step could be found in my Jupyter [notebook](notebooks/Japanese Food Classification.ipynb). Because of the way I split my data, I only had around 300+ images per class for training, which isn't a whole lot. To mitigate this problem, I used Keras's ImageDataGenerator for image augmentation. This allowed me to augment the data on the fly and get better training results.
I attempted several models include a base ConvNet model and other transfer learning models. For now, the best model I came down to is to use MobileNet transfer learning as a feature extraction. I fit my model with a RMSProp(lr=0.0002) optimizer first, and then fine-tuned it with a lower learning rate. (Note: I am still working on getting better accuracies).
With my current best model, I created a flask web application in the web_app folder that could be run with the code shown above. If you have another model that you'd like to test out, you can move your best model into the models folder inside web_app and run it there.
- Clean current data (remove unwanted images)
- Unfortunately, as the images were scraped off from a search engine, not all the images were of good quality. Some images were just the storefront that sold the dish, some were advertisements, and some were even plastic packagings of the food.
- Get more data (images) for the current classes
- Train a model for another cuisine