Skip to content

Latest commit

 

History

History
33 lines (24 loc) · 2.09 KB

File metadata and controls

33 lines (24 loc) · 2.09 KB

ImageProcessingHackathon-SuperAI

By: Parinthapat Pengpun

This repo holds winning 1st Place Notebook in the SuperAI Season 3 Image Processing Hackathon!

What is the competition?

This Image Processing Hackathon is one of the 6 hackathons for qualifying entry into the Super AI Season 3 Program. This hackathon is open to the public meaning that people of all ages and all levels of education all over the nation are participating.

The goal of this hackathon is to classify digits (0-9) from images which could be written, printed, or from anywhere in the real world. The dataset is very diverse with the digits varying in size, color, and contrast.

Composition of Project

  • Image Processing Techniques
  • Usage of State-of-The-Art Vision Transformer Model

Rationale behind Image Processing Steps

Therefore, I tackled this by making all the images as similar as possible.

  1. Resize every image to fit 224x224 pixels (keeping aspect ratio) and pad the remaining spaces.
  2. Apply auto contrast to increase the contrast between the digits and the background
  3. Change the image to grayscale in order to emphasize the contrast and reduce classification based on digit color
  4. Save the image.
  5. Invert the grayscale image and save it as another image (effectively augmenting the dataset) to handle for black and white text.

Choosing a Model

The final aspect of this was also to choose a SOTA model. After reading a few papers, I settled on using Vision Transformers due to the nature of their architecture.

Code

You can access the code under the Image_Processing_Hackathon.ipynb file.

Special Thanks

All in all, this was a very informative experience and I have to thank AI Builders for providing me with the knowledge to do all of these things!