# Convolution with a non-square kernel

So far I've only encountered convolution kernels which are square (ie, have the same rows as columns).

Are there any cases in which a non-square kernel makes sense? If not, why?

• 2d convolution is a general operation that one encounters under many circumstances. The answers below focus on convnets for image processing, but what situation are you actually asking about? – user20160 Jun 13 '18 at 8:49
• Square kernels fit nicely within matrices. One could see any square convolution with structural zeros as non square. A kernel detecting a cross like shape would be an example. (See the wikipedia kernel examples page.) – spdrnl Jun 13 '18 at 8:55
• @user20160 The answerers guessed correctly. I just added the image-processing tag. – Tom Hale Jun 15 '18 at 8:53
• @spdrnl did you mean this page? – Tom Hale Jun 15 '18 at 8:56

For concrete example see this network. It is applies $4 \times 1$ kernels to short-time Fourier transform of sound (it runs on $513 \times 128$ input).