I have a neural network that I trained on 32 * 32 px size images. Can I use these filters learned from the network on larger images not used in training the network such as a 600 * 800 px image? Or does it not make any sense to apply filters that were learned on smaller images on larger images and vice versa?

Here is an example of some images: An image from training set

32 * 32

Here is a random image that I got online to test 480 * 360 (not really 400*600)

Here is when I applied the filter(s) Result of passing a filter

  • $\begingroup$ If you give me sample images, both the 32x32 and the 600x800 then I might be able to give R-code in my answer. I'm assuming they are grayscale. $\endgroup$ Apr 30 '16 at 23:25

Your question:
Can I use convolutional filters used on 32x32 px size images upon 600x800px images?

My answer:
If the features are the same sizes (aproximately) then zero pad the edges and go to town. If not, then there are tricks you can do that will allow them to be used with slightly more noisy results.

Options include:

  • sub-sample the "big" pictures so they are on the scale of the
    learning. Convolutional methods work on sub-pixel resolutions. Sub-sampling can be "pick random", "pick mean", "pick median" or other. Personally I like to pick about 30% of the way between the mean and the median, but it is just personal preference.
  • interpolate rows and columns in the convolutional templates, so that they operate as if they were larger. Zero padding rows/columns can also work

The convolutional templates are invariant on translation, not the other operations (rotation, skew, scale).

I think there are higher-dimensional analogies to the FFT, cousins and second cousins, that are invariant in those other operations, but I do not know what they are and they are higher dimensional. From my "mathematical modeling" undergraduate course sponsored by Lockheed, I am sure that whatever they are - they are of interest for folks who perform extensive photogrammetry. Those analogies would work seamlessly without having to account for rotation, skew or scale.


Your Answer

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

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