1 of 8

Coloring Butterfly Images using GANs

By Jordan Conragan, Ryan Cummings,

Adam Goldstein, and Priyanka Moorthy

2 of 8

Project Objectives

Butterfly Colorization:

  • Color grayscale images of butterflies

​

Butterfly Interpolation:

  • Recreate a butterfly image from its line art

Why butterflies?

  • There are many butterfly datasets available

​

  • Butterflies are colorful

3 of 8

Preprocessing

Image Segmentation

Grayscale

Edge Detection

Extract

Butterfly

4 of 8

Generative Adversarial Networks

  • Trained a GAN to generate synthetic butterfly images
  • Generator function used to create potential plausible butterfly images
  • Discriminator function used to classify these examples as real or fake
  • Attempted with both grayscale butterfly input and lineart input

5 of 8

Butterfly Colorization

6 of 8

Butterfly Line Art Interpolation

  • Condition on input images(line art) and generate corresponding output images(butterflies)

Generator Loss

gan_loss + LAMBDA * l1_loss

Discriminator Loss

real_loss + generated_loss

7 of 8

Conclusion

All project goals were met

​

​

These are some of our favorite results

8 of 8

just pasting that here if anyone wants to use it

Pix2Pix cGAN outputs