anomalous_drawings
Flask app that detects anomalous bowtie drawings: https://anomaly-bowtie.herokuapp.com (this link might load a bit slowly when you click it the first time)
This app uses a simple autoencoder to perform anomaly detection on bowtie drawings. You can draw a bowtie (or anything) in the left canvas and the right canvas will show you the autoencoder's reconstruction (784-32-782 model --> 28x28 image). If your drawing is not a standard-looking bowtie (anomaly), the reconstruction canvas (right) will turn red. A good reconstruction of your drawing will give a low cost (not anomaly) while a bad reconstruction will give a high cost (anomaly). The anomaly threshold is set at 0.07: error > 0.07 --> anomaly.
The flask app resides in the app submodule and is hosted on heroku. This app is responsive and has touch support (tested in Chrome).
Data: The bowtie drawings used to train the autoencoder are a subset of Google's Quick, Draw! dataset.
Notebook that explains how anomaly detection can be performed via an autoencoder: Bowtie Drawing Autoencoder
Notebooks about the autoencoder's implementation can be found in my ML algorithms repository. model.py and utils.py are based on this repository's notebooks.
- Since Google's Quick, Draw! (preprocessed) dataset resembles the MNIST dataset in its simplicity, I decided to reuse the autoencoder architecture I used for the MNIST dataset on Google's dataset.
I used this HTML Canvas tutorial in my flask app: https://dev.opera.com/articles/html5-canvas-painting/