An Android application to classify ingredients, then fetch recipes using the classification.

If the original classification is incorrect, the user has the option to select a class out of the Top-5 highest probability predictions.

With there being multiple compontents of the app (the CNN being trained, the server used to host the CNN for classification, the Android app itself), I've tried to separate each into its own sub-directory:

Python Code - Contains the sub-directories Flask Server & PyTorch CNN Training, with the component's corresponding scripts.

MobileRecipeGenerator - Contains the Android Studio project, seen within the demo provided above.

When training the neural network, I used the Fruits 360 dataset.

With the size of the dataset being so large, I decided to exclude it from the repository, but can be found here.

I also removed a large number of classes from the dataset (62 classes remain) as it contained objects not commonly seen in UK supermarkets - the class list can be seen under Python Code > FlaskServer > app.py > class_types.

The CNN is hosted on a server, where the android app calls an endpoint.

The main reasoning behind this was to reduce the overall size of the app, whilst increasing classification speed (by it being ran on a pc with better specs).

To run the server locally, simply start the app.py script.

The android app has been tested on multiple emulators within Android Studio (multiple Android versions), as well as on a real device (Samsung Galaxy S9 Edge) - as seen above.

To connect to the server, the constant ** within Helper.java needs to be changed to the ipv4 address of the device running the Flask server (typing ipconfig in a cmd session).

Anaconda was used to install dependencies needed, the packages used are:

Python 3.9

• torch
• torchvision
• Flask
• numpy