0

I would like to try to implement a autoencoder fully-connected convolutional neural network as Unet to transform an image into another with an unknown non-linear relation between them.

I've got Gaussian kernel convolution algorithm which works well but I would like to try something with machine learning approach.

Have you got any idea of an other ANN architecture? without standard fully-connected ANN (already tried with good results) Thanks a lot

Menol
  • 1,230
  • 21
  • 35
lucasdbo
  • 1
  • 1
  • Hey lucas, can you post your architecture? And are you working with a segmentation problem or you want to reconstruct your image with some small changes? – Filipe Lauar Jan 16 '20 at 14:59
  • Hi Filipe, I tried FC ANN with 11 features (position in matrix, pixel intensity and intensity of 8-connect neighbors) and one hidden layer. it's not a segmentation problem. Yes i want to reconstruct my image with some small changes as blurring effects. – lucasdbo Jan 16 '20 at 15:13
  • Hmm, to reconstruct the image with small changes it's better to use a Variational Autoencoder (VAE). I'm gonna do an answer to explain how does it works. – Filipe Lauar Jan 16 '20 at 17:03

1 Answers1

0

as you want to reconstruct your input image with small changes, Variational Autoencoder (VAE) is a better architecture. The VAE works just like an Autoencoder, but its bottleneck is a probability distribution with mean and standard deviation. You can create new images with small changes from the input image changing the latent space (bottleneck). This is an image with a VAE architecture:

enter image description here

If you have any doubt feel free to ask.

Community
  • 1
  • 1
Filipe Lauar
  • 434
  • 3
  • 8