Project in a nutshell
The purpose of this project was to use a machine learning platform to train a model to accurately tag graphics that are uploaded to Pixel Scrapper, which is a digital scrapbooking website.
Our Process
The tags are based on a database of already tagged graphics on the Pixel Scrapper site. Many of the graphics are abstract art (meaning that they don’t have concrete objects that are easily identifiable), which is why an existing model could not be used. We investigated different cloud ML platforms, such as Google AutoML, Microsoft Azure, and AWS SageMaker, to use for our project. We ended up choosing AWS SageMaker, due to their competitive pricing and the fact that Pixel Scrapper's data was already in AWS. The primary technical tasks that we completed were image processing, training and tuning our model, and deploying and testing the model.
How the project will be used
The auto-tagging will initially be used by the Pixel Scrapper team of designers, which are people uploading commercial use graphics to the site. This will help speed up the process for when they do bulk uploading, because they won’t have to manually tag each one of their graphics. Pixel Scrapper may also release auto-tagging to users down the road; this would be helpful for users because some of them have a hard time accurately tagging their graphics. It’s crucial that the tagging is accurate, so that people can go on the site and use tags to find different kinds of graphics that they want.
We would like to acknowledge our project partner Jordan Magnuson for his guidance on this project.
| Attachment | Size |
|---|---|
 Team milestone draft | 89.29 KB |
