Intercepting convolutional kernel inputs

I was wondering if it is possible to somehow grab the inputs NxN that are processed for each of the kernel convolution on each pixel of the image?

I’m trying to analyse each of the kernel data before and somehow connect it with a specific output neuron.

Thank you.

I’m not sure, if I understand your use case correctly, but if you want to compute the convolution manually, you could use unfold. Have a look at this post, which might be a good starter.

Thank you very much for the useful info!