Combining images and numerical values in a deep neural network

25 vues (au cours des 30 derniers jours)
Luca Ghilardi
Luca Ghilardi le 2 Juil 2019
Commenté : shlomo odem le 14 Nov 2021
I want to combine in a encoder/decoder architecture (like U-Net or FCN) for semantic segmentation both images and numeric values. The images are aerial maps from which I will take random crops to feed the convolutional part of the network, I would also like to add some data in the feature space (like sun inclination) by connecting the fully connected layer of the network with one or more neurons for the numerical part.
0*Umo-FPkUXDRgXK0L
My doubts are:
should I use a single input layer that accepts both types of data?
If not, how I can handle the training and datasets with two different input layers?

Réponse acceptée

Gabija Marsalkaite
Gabija Marsalkaite le 4 Juil 2019
Probably simplest way is to add other features as channels. There is example by Uber Engineering where they added gradients for x and y to ease coordinate learning:
https://eng.uber.com/coordconv/
Right now (2019a) Deep Learning Toolbox does not support multi-input layers but we are looking at this and this may change in future releases.
  1 commentaire
shlomo odem
shlomo odem le 14 Nov 2021
i have the same problem, did 2021a support multi-input layers?
how can i do this network?

Connectez-vous pour commenter.

Plus de réponses (0)

Catégories

En savoir plus sur Image Data Workflows dans Help Center et File Exchange

Produits


Version

R2019a

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by