I recently read Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Network by Wenzhe Shi et al. I cannot understand the difference between deconvolution (mentioned in Section 2.1) and the Efficient sub-pixel convolution layer (ESCL for short) (Section 2.2)

Section 2.2 defines the ESCL as performing convolution with a fractional stride. Isn't this what deconvolution is? Why is a distinction being made?


  • $\begingroup$ Theory is good. But in practice I getting much clear result with sub-pixel conv rather than deconv, also sub-pixel learns faster. $\endgroup$
    – iperov
    Commented Dec 31, 2018 at 8:09

1 Answer 1


In this paper, theauthors have provided a deep explanation, basically its the exact same thing. sub-pixel uses or^2 conv kernels all at res H/r,W/r while the regular transpose conv uses o conv kernels all at res H,W the authors prove that one can rearrange the feature map (Or^2,h,w) to a feature map (O,hr,wr) without losing information.


Your Answer

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

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