lyesds/faktion-usecase

Anomaly detection in pictures

Computer vision for anomaly detection in pictures

Project • Data source and structure • Our solutions • How to use

The project

The aim of this (learning) project is to classify pictures of dice as anomalous vs. normal:

The dice on the left are anomalous while those on the right are normal.

Data source and structure

The images used to build this project have been provided by Faktion and are available here. Please grab these assets and extract them in the root directory.

.
├── assets
|   ├── anomalous_dice
|   └── normal_dice
├── utils
├── .gitignore
├── main.py
└── README.md

Our solutions

1. OpenCV manipulations

This first approach consists in comparing any input picture of a die to be classified as anomalous vs. normal with templates of normal dice. If the input die is normal it should show a difference with the templates that is close to "zero". Whereas if the input die is anomalous, there should be a significant residual difference with the templates.

We found that the difference - in terms of pixels values - is a fair proxy of the dice class (normal or not) and allows to reach a 0.84 F1-score for the anomalous class on the original dataset.

2. Convolutional neural network (CNN)

This second approach consists in ...

How to use

You'll need Python installed on your computer to clone and run this application. From your command line:

# Clone this repository
$ git clone https://github.com/hakanErgin/faktion-usecase

# Go into the repository
$ cd faktion-usecase

# Install dependencies
$ pip install requirements.txt

# Run the main.py script
$ python run main.py

---

GitHub @hakanErgin @lyesds @nicesoul