1
$\begingroup$

I try to be brief. For what I understood Convolutional Neural Network CNN for style-transfer extract/learn the main features of source Image. I think these features basically are Kernels (of different sizes depending on Levels). What I ask is how to reconstruct image once we obtained the Kernels. To simplify let's assume I have N-kernels of the same size. How should I reconstruct the Image using only the Kernels ? If question is not clear I'll try to clarify it. (I try using and weighting their responses, but not succeed) Thanks.

PS I wrongly posted to stackoverflow the question... but I think this DSP is the right place. link to Stackoverflow: https://stackoverflow.com/questions/41833641/image-reconstruction-using-kernels

$\endgroup$
1
$\begingroup$

I think what you would like to do is something similar to an Auto-Encoder, in which one would append an Encoder to reconstruct back the image. The 'kernel', or more appropriately the intermediary representation is used to synthesize an image, and an appropriate reconstruction objective is minimized to recover the input. On the other hand, you could always train a second upsampling network, which reconstructs the input image, given a 1D/2D feature.

$\endgroup$
  • $\begingroup$ my goal is to obtain a basic "style-transfer" such as CNN-Style-Transfer. I've already implemented an algorithm to obtain a specified number of kernels that are the so called "features" of the input image. Basically they are the K-mean of patches of the input pictures. I'd like to reconstruct the image using these kernels. And, most interesting, reconstruct a different input image using the features. Some CNN Kernels example image Thank you for the reply. $\endgroup$ – reexre Feb 6 '17 at 20:12

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.