I am trying to understand if Semantic segmentation with U-NET. Are we training kernels to extract features or are we training a fully connected layer at the end? Or both? Because based on this image:
I don't think there is a need for a fully connected layer at the end. If we just train our kernels, U-net will just do the image segmentation. I couldn't find any resource for "how to train unet model for image segmentation" so i wanted to ask here. Thanks in advance
Asked
Active
Viewed 520 times
-1

Christoph Rackwitz
- 11,317
- 4
- 27
- 36

canovich
- 1
- 4
1 Answers
1
U-Net is a "fully convolutional network" (CNN). These types of architectures do not have any fully connected layers in them.
Consequently, they do not depend on the input image size - and the size of the prediction varies according to the input image size.

Shai
- 111,146
- 38
- 238
- 371
-
Thanks for your answer. So what does the last layer of U-net doing? I mean why do we need the last 1x1 convolution? Other convolutions kernels can already be trained for image segmentation and can differentiate images for segmentation(like in the picture that i added on my question). Am i wrong? – canovich Jan 20 '22 at 13:20