Convolution on single image with 3x3 filter

vision
1

Hi, I am trying to make something like sobel filter on a single image in PyTorch.
I did that in Pyton using CV2.

image
image613×790 144 KB

How to do that for a single image in PyTorch?

2

Hi,
You need to use torch.nn.functional.conv2d in order to convolve the image with an specific kernel.
docs indicates the shape of the kernel, very straight forward.

Note that pytorch use cross-correlation instead of convolutions.

Regards

2 Likes
3

Yes, JuanFMontesinos, I used just what you told me.

image
Where f is sobel filter expanded to all 3 channels.

import torch
filter = torch.tensor([[1., 2. , 1.], [0., 0., 0.], [-1., -2. , -1.]])
f = filter.expand(1,3,3,3)
1 Like
4

Wonderful feedback. Thanks.