6
$\begingroup$

I am implementing a paper on image segmentation. It is based on the slight modification of the u-net architecture.
The paper is based on encoder and decoder steps Following are the lines of the paper which I am unable to understand.

In the decoder path, features from the encoder path at the same 
resolution are fused through concatenation.

and then the architecture of the network is given-

Layer .  Output size .  Filter Size    Stride Dropout
Inputs .  _ .              _ .           _ .    _
Conv1
Conv2
Max Pool1
Conv3
Conv4
Max Pool2
Conv5
Conv6
Max Pool3
Conv7
Conv8
Max Pool4
Conv9
Conv10
Upsampling1
Conv11
Conv12
Upsampling2
Conv11
Conv12
Upsampling3
Conv11
Conv12
Upsampling4
Conv11
Conv12
Output

I am not adding the other columns values for conciseness. Anyone who want to read the whole paper, it is this. The required lines are in the Table 1.

$\endgroup$
6
$\begingroup$

Generally, if you look at image segmentation models, they have two main paths, what the author of your paper calls encoder and decoder paths.

The role of the encoder is to contract the size of the image while extracting meaningful information, while the decoder restores the contracted image to its original dimensions. However, a lot of information is lost during contraction. To overcome this issue the decoder takes some features directly from the contraction path via skip connections.

A certain decoder layer's input is actually the concatenation of two things:

  • The output of its previous layer
  • The output of the respective layer (the one with the same dimensions) of the encoder

This becomes more clear if you look at the following image:

The left part of the network is the encoder and the right is the decoder. The skip connections are the grey horizontal lines that go directly from the encoder to the decoder.

These can't appear the way you portray the model because you show only the layers and not any information about how they are connected. I think you assume the network should be sequential, but segmentation models usually aren't.

$\endgroup$

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.