GrayScale to Color conversion

Table of contents

• The challenge
  • Overview
• My process
  • Built with
  • Result of classifier
  • Useful resources
• Author

The challenge

Create a program to determine the understanding of U-Net Learner, resnet34 and deep learning skills. Handson experience on Jupyter Notebook code.

Overview

In this project, we have used U-Net learner for converting grayscale black and white images into colored ones.

My process

• First 200k Colored Human Face images Dataset was taken from Kaggle.
• Then a new dataset of grascale images were created by converted all images into grayscale using crappifier method.
• Both dataset were trained using U-Net learner with resnet34 CNN model. The Cost funtion used was MSELossFlat().
• The training was done on Google Colab with GPU Hardware accelerator. It took around 3 hours and 20 minutes to complete the training.
• It ended up with a 6% Validation error.

• Testing on the existing data.

• You can save your model in .pkl file, which can be integrated in other webapps or mobile applications.
• Below is the link for .pkl file for trained model.
• PKL file link

Built with

• Python 3
• pandas
• fastai

Result of classifier

Useful resources

• CelebFaces Attributes (CelebA) Dataset
• Fastai Documentation - This helped me for understanding all concepts of fastai library. This is the top of my list for machine learning development.

Author

• Abhinav Shukla