Colorizing images with Auto Encoders
Auto Encoders (AEs) enable us to produce a dense and low-dimension representation of our samples by removing noise and preserving the main concepts of the samples (1). Although, it is the main purpose of them, we can abuse them for predicting some thing different than the input. Denoising Auto Encoders (DAEs) are a very well-known example of this. It turns out that with a little bit creativity, the simple idea of AEs can be used to create interesting applications. In this post I am going to share my experience with AEs to colorize black and white images. The idea is to first create a dataset of images in which each image is paired with its black and white version. Then input the black and white images to AEs and put the colored images at the end, forcing the model to create three-channel output (in RGB color space). You can see the code here (github) or here (colab).
- You can find my playlist (Introduction to AEs) here (in Persian).
