1
$\begingroup$

When an image is fed into a CNN, it would pass through different layer of different filters. The visualization of these filters look like:

enter image description here

enter image description here

Here we can safely claim that the 1st filter is more activated than the 2nd hence it carry more weight in identifying the input image. My questions is, how do we quantitatively compute the "activation" (or the usefulness) of a filter?

I have tried 3 approaches: by brightness, by variance and by Shannon entropy. Yet none of them working perfectly so far.

When measure them by brightness, I got upper image brighter than the lower image:

enter image description here

enter image description here

When measure them by Shannon entropy, I got upper image with higher entropy than the lower image:

enter image description here

enter image description here

So, any suggestion how could I measure their's activation?

$\endgroup$

0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.