-1

im working on training a model to colorise black and white images. so far, ive found 2 methods;

  1. transfer learning using vgg16
  2. autoencoder from scratch

ive did both and noticed transfer learning using vgg16, although it trains faster, it produce awful results as compared to implementing an autoencoder. why is that? Also, can you provide more methods?

desertnaut
  • 57,590
  • 26
  • 140
  • 166
Vortekus
  • 69
  • 1
  • 9
  • I’m voting to close this question because it is not about programming as defined in the [help] but about ML theory and/or methodology - please see the intro and NOTE in https://stackoverflow.com/tags/machine-learning/info – desertnaut Jul 30 '22 at 23:37

1 Answers1

0

VGG-16 with ImageNet weights might not be suitable for a colorisation problem. VGG is a neural network developed/trained for Image classification problems. You can use ImageNet weights that'll help in transfer learning for custom image classes like Cats vs Dogs or Cars vs Bikes (Commonly found classes).

Auto-encoders are a different class of neural networks altogether that focus on input -> output mappings. An auto-encoder is perfectly suited for a problem like this, you input one type of image and get a modified version of the same image, be it colorisation or denonising to name a few uses of Auto-encoders.

As an alternative, you can use U-Net in combination with GANs to tackle this problem statement. More details here

Pranav Raikote
  • 97
  • 1
  • 1
  • 6
  • can you provide more information or link resources to why autoencoders is superior to vgg16 in this context thank yo – Vortekus Jul 31 '22 at 08:21
  • 1
    VGG is a neural network developed/trained for Image classification problems. You can use ImageNet weights that'll help in transfer learning for custom image classes like Cats vs Dogs or Cars vs Bikes (Commonly found classes). Auto-encoders are a different class of neural networks altogether that focus on input -> output mappings. – Pranav Raikote Jul 31 '22 at 11:28
  • 1
    Adding a clarification with a "comment" that is almost as long as the original answer does not make much sense - please **edit & update** your answer accordingly. – desertnaut Aug 02 '22 at 09:41