2
$\begingroup$

I am trying to understand how the deconvolution works in Convolutional Neural Network for image segmentation problem. I have seen this definition:

Filters used in deconvolution is just the transpose of the convolution matrix.Figure from CS231 videos.

Does this mean that the convolution layer and deconvolution layer shares the same filters? Do we train a separate set of filters for deconvolution layer?

$\endgroup$

3 Answers 3

2
$\begingroup$

Here, I think, you can find good visuals and explanations for convolution/deconvolution arithmetic.

$\endgroup$
1
  • $\begingroup$ Thank you for your prompt answer. I have looked through the material. It is very useful and clear but I still have a little confusion. It looks to me that 1. In convolution layer, we reshape our filters to form a matrix (w) so we can do matrix multiplication; 2. In deconvolution layer, we take the transpose of the matrix (w from convolution layer) and take that as the set of filters to use in deconvolution. Is this correct? $\endgroup$
    – Nougat Xu
    Oct 3, 2018 at 20:38
0
$\begingroup$

Here's a very nice intuitive explanation on why we use the same set of filters in both convolution and deconvolution layer.

$\endgroup$
0
$\begingroup$

No, these two layers do not share the same filter parameters. By coding and decoding you increase the representation power and enlarge the receptive fields.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

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