Create a MNIST-like data set for Mahjong using Scikit-image

By Cheuk Ting Ho

Elevator Pitch

I play Mahjong, a game popular in East Asia. There are fixed tile designs of each Mahjong tile so it should be easy to detect what tile it is using machine learning. But first, we need a big enough data set for training. We will find out how to generate > 1000 images from 1 image in this talk.

Description

Goal

Audiences can learn from my experience what to do if they do not have enough images for their computer vision project. Also learning what Scikit-image provide and how to use it.

Target Audience

Knowledge of Scikit-image and computer vision is not required. However, it is assumed that the audience have basic concept of how machine learning works and have some experience using Scikit -learn.

Outline

  • Why do I want Mahjong tile detection?
  • What do I need in the project
  • Using Scikit-image to generate more images for training
  • Testing with Scikit-learn SVM
  • The rest of the project and what I want to achieve